Patent Grant
8132190
This invention relates to a management method, more particularly by recording or external monitoring, for software operations carried out during the execution of an application process executed on a computer. This management applies in particular to the recording of operations in the form of logging data, or to a replay of operations according to a given execution corresponding to a previously logged execution.
The method is particularly advantageous applied to operations performing the internal events of this process, which can itself be part of a multi-process and/or multi-computer software application.
The invention also relates to a system implementing such a method in the functioning management of the software applications that it is executing.
The field of the invention is that of computers, isolated or in networks or “clusters” of computers formed from a number of computers working together. These clusters are used to execute software applications providing one or more services to users. Such an application can be single or multi-process, and be executed on a single computer or distributed over a number of computers, for example as a distributed application of the MPI (“Message Passing Interface”) type or “Shared Memory” type.
In particular, the invention enables to carry out a management of the functioning, within the cluster, of such an application termed master or primary, for example by another software application termed intermediate application, for example an application of the “middleware” type. This functioning management may comprise, inter alia, the operations of replication, redistribution, reliabilization, or tracing or debugging of all or some of this application, within the primary node or in collaboration with other nodes termed secondary.
In the context of this operational management, it is often useful to log the functioning of the primary application or of one of its processes, i.e. to record the data representing this functioning, enabling the execution to be reconstituted. Along with the primary application execution, this data is then generated in the form of logging data and is transmitted to one or more secondary nodes for storage and backup.
For example in order to trace and study the functioning of the primary application in detail, it is then possible to study or to reconstitute this functioning, later on or remotely, in a controlled and monitored manner.
Also as an example, if the primary application experiences a failure, in particular a hardware failure, it is then possible to create a new standby application on a secondary node in order to replace the services provided by the primary application. This standby application can then be created in a known state, for example a restart point state recorded previously. From the logging data of the primary application, it is then possible to force the standby application to reconstitute the execution of the primary application up to the time of the failure. After this reconstitution, or replay, the standby application is in the same state as the application until the last event, the logging data of which have been received outside the primary node. If all the events preceding the failure have been logged and transmitted up to the failure, the standby application can then take over with little or no interruption of the service for the users.
However currently, many existing applications do not have such management functionalities, and it would be too complex and costly to modify them in order to add these to them.
The solution which consists of implementing these functionalities in the system software of the computer or of the primary node presents some considerable drawbacks, such as the risk of errors, instability or incompatibility within the network and the requirement for special skills in the field of systems software.
In addition a solution is proposed by the authors of this invention, which consists of these management functionalities being taken over by an intermediate application which is mainly executed in the user memory space and requires only a few modifications within the system software itself.
However, in this type of solution, inter alia, the creation and processing of logging data, as well as its the transmission from the primary node to a secondary node represents a significant calculation load with respect to the execution of the primary application itself, as well as for the communication networks used. In the prior art, the master application then experiences such a loss of performance that, often, this functioning management cannot be satisfactorily used in exploitation situation.
In fact, in order to be able to represent in a coherent manner, or even a complete manner, the execution of the primary application, the events to be recorded and to be transmitted are often very numerous. Moreover, the majority of these events correspond to operations the execution of which is very fast, in particular the events which are internal to the hardware or software resources of the primary node, for example a system call requesting the assignment of a semaphore or reading an item of data in memory.
By contrast, for each of these events, the generation and storage, as well as the transmission of the logging data is a much longer operation, in particular for the internal events.
In fact, logging each event is in itself a process which requires at least one and frequently a number of software operations, each of which constitutes a load and a working time at least equal to the logged event in itself. According to the implementations and the type of internal event, the logging adds for each event an increased load or working time larger by a factor which commonly ranges between 100 and 10,000.
Furthermore, the hardware and software protocols used for transmission to the outside of a computer have performances which are in general poor in relation to the number of events logged, which is also a disturbance to the use of the network as well as a bottleneck for the performances of the master application.
Certain solutions exist which allow the number of events to be logged to be reduced, in particular by not logging events of a non-deterministic type.
An event, or the operation which constitutes it, in particular a software operation, can be qualified as deterministic if the result of its execution depends only on the initial conditions which existed at the time of its initiation. More particularly, in the context of managing a unitary operation or an execution or a functioning as described here, an operation is termed deterministic if it is deterministic from the point of view of the process which initiated it, i.e. if the result which it sends to this process depends only on the initial state of this process. Similarly, a contiguous succession of deterministic operations may constitute itself a deterministic sequence.
Within the running of an application process, many of the operations performed are deterministic, in particular among internal operations. For example, mathematical or logical internal operations will be deterministic more often than not if they affect only those resources forming part of the initial state of this process, and which it alone can modify.
A contrario, some operations applying to shared resources are often non-deterministic vis-à-vis such a process. For example, a request for attribution of a shared semaphore or a “lock” covering a memory zone shared with other processes could be non-deterministic. In fact, the result, i.e. the obtaining or otherwise of this lock or this attribution might depend on the state or actions of other processes, which would sometimes have reserved or not reserved this resource.
However, the replay and, in particular, the logging of non-deterministic events still constitute a loss of performance which could usefully be reduced. In particular while running the master application, the logging operations represent a work load for the operational node, and can be the cause of a fall-off in performance due to the action of the intermediate application.
One aim of the invention is to overcome all or some of these drawbacks.
The invention proposes a method for managing a non-deterministic software operation initiated by a program instruction within the unitary operation of an application process, termed managed process, executed by at least one computer, in the case where the performance of said operation sends to said managed process at least one result datum, termed actual result.
This method comprises the following steps:
The invention therefore proposes to perform a management of non-deterministic operations which depends on their conformity or not to a known or calculable deterministic prediction, capable of being viewed as a sort of reproducible compliance grid. It is therefore possible to reduce the management workload in cases where the result of an operation conforms to this grid.
This gain is particularly sensitive as regards logging operations, which must be recorded not only in the primary node, but also often transmitted to the outside, and this during the exploitation of the primary application.
The invention thus produces a form of predictive or heuristic compression of the operations or events to be logged, and can reduce the proportion of these operations or events which require integral management. In fact, if the operation behaves as predicted, the result which it sends no longer needs to be logged in itself as its value can be found later according to a known method which could be re-used at the appropriate moment.
The invention in particular proposes to manage an operation, termed logged, included in an execution sequence of a logged process executed in a computer, termed primary node, this management recording said logged operation in the form of logging data enabling a restart process executed by a secondary node to replay an operation, termed replayed, corresponding to the logged operation and sending to the restart process a result corresponding to that of the logged operation.
In the case of an actual result not corresponding to the expected result, i.e. an unpredicted operation, the complementary management phase then comprises a storage of the logging data, including a result datum representing the value of the actual result for the logged operation.
It is therefore possible to reconstitute the exact execution of the sequence during a future use of this logging data in the case of an unpredicted result.
During the logging of a sequence, the phase to test the operations to determine their deterministic nature or non-deterministic nature represents an additional load. However, from a certain proportion of operations predicted correctly within this sequence, it is clear that the savings achieved by not logging the predicted results can become significantly larger than the overload due to this test.
In the opposite case, i.e. a predicted operation for which the actual result corresponds to the expected result, the invention enables to avoid storing and/or sending to the outside the data from the result of the predicted operation, which constitutes a workload gain.
Optionally, it is possible to provide storage for logging data which includes an identification datum for the logged operation, for example in the form of a sequencing number. This identification datum can keep a trace of the execution of this operation in order to be able to include it in the future use of this logging data.
As the prediction function is deterministic and known, the result can be recalculated with certainty during the replay, from the initial conditions of the operation. If the result during the replay does not conform to this same prediction, the result of this prediction could be used to “force” the result of the operation such that the restart process actually receives the same result as during the logging.
According to the invention, a deterministic operation may be logged without storing or transmitting the result of the operation, which is a saving in terms of workload for the primary node.
More particularly, the method can record a sequence, termed logged sequence, of logged process operations, said sequence including at least one operation, termed unpredicted, sending an actual result not corresponding to the predicted result. This recording then comprises storage of at least one log file enabling the restart process to replay a sequence, termed replayed, of operations corresponding to the logged sequence operations. In this context, said log file comprises data representing the actual result, termed logged, of said unpredicted operation and which can be used so that, on completion of a replayed operation corresponding to said unpredicted operation, the restart process includes a result corresponding to said logged result.
According to one embodiment, during the logging of a logged process sequence, at least one predicted and non-deterministic operation is logged without being stored within the log (JSem1) of the logging data representing the actual result (DR) of said predicted and non-deterministic logged operation (EVInD), by incrementing a identification datum corresponding to this predicted operation.
Optionally, the incrementing sequence can be reset during the complementary management phase carried out on an unpredicted operation.
The logging data stored in the log for logging at least one unpredicted operation can thus comprise:
Advantageously, the invention proposes in particular to record within a log at least one logged sequence, this recording comprising an iterative repetition, for each logged operation constituting a non-deterministic internal event of at least one type within said sequence, the following steps:
Advantageously, only unpredicted results give rise to storage of the result.
According to one variation, the sequencing datum may only be incremented on an operation with a predicted result, and be reset during an operation with an unpredicted result.
In parallel with or independently of the logging, the method according to the invention enables managing a sequence, termed replayed, of operations forming part of a restart process, executed in a computer termed secondary node.
By using a replay prediction function which is identical or corresponds to that of logging, the invention thus produces a form of predictive or heuristic compression/decompression without data loss, which can provide during the replay the same results as during the logging, together with reducing the amount of logging data to be stored and transmitted. The logging data can thus be used to perform a replay of the logged operations in the form of operations replayed by a restart process.
This restart process may, for example, be launched and then executed by itself from its executable file(s). During the execution of this replay, the replayed operations thus provide themselves a result each time that their character orders this.
The management according to the invention may thus use the logging data to procure that, in each case of a non-deterministic replayed operation, the restart process takes into account a result, termed forced, corresponding to the result returned during logging.
According to the invention, the complementary management phase thus comprises, for at least one unpredicted non-deterministic operation at the logging time:
Therefore, each time that the restart process executes, within its executable file, an instruction performing an unpredicted operation during the logging, the complementary management phase enables then to check or force the result in order that it conforms to the logging data, and thus produce a execution of the restart process conforming to that of the logged process.
On the other hand, when the result of the operation has been predicted correctly during logging, the logging data may contain no result datum, even if the operation is non-deterministic and its actual result uncertain.
For at least one non-deterministic operation which has been predicted during logging, the complementary management phase may therefore comprise the following steps:
It is therefore possible to find the actual result obtained during logging even if it has not been stored, and to check that the replayed result conforms to it, or to force it if it does not conform.
In addition, for each operation for which the logging data does not comprise the logged result, the complementary management phase includes an incrementing of the sequencing value representing said replayed operation.
Therefore, the incrementing enables to monitor the progress of the restart process, without intervening in its running as long as it is neither useful nor necessary to maintain a compliant running. The fact of not intervening more than necessary is therefore a saving of work for the process or for the application managing the replay, or for the computer which executes them.
In particular, in the case of a sequence comprising both predicted and unpredicted operations, it is possible to obtain a replay which corresponds to the execution of the logged process, or even is exactly identical to it, together with using logging data of a reduced volume and the storage of which requires a lower workload.
The method according to the invention may be implemented in particular to manage all the operations producing, within the managed sequence, events of at least one type.
It is thus possible to perform a logging or a replay which represents all of the execution of the process vis-à-vis certain particular items, for example another process, a user, or a particular resource.
Furthermore, this method may be used to manage all the operations identified as being intrinsically non-deterministic or sending a result identified as being intrinsically non-deterministic.
In various versions, this method of heuristic optimisation may be combined in different ways with other optimisation methods.
For example, within a solution where only deterministic operations are not logged, it is thus possible to optimise the workload even more by logging only part of the remaining operations, i.e. only operations that are both non-deterministic and unpredicted.
More particularly, the method according to the invention manages all the operations producing, within the managed sequence, non-deterministic events internal to the managed process, or to its application, or to the computer which is executing them.
In coordination with a management of external events, the invention can therefore manage the entire execution of the process, in particular to enable a logging during exploitation, at the same limiting any slowing down of the managed application.
Such a management, for example logging or replay, is advantageously applied to at least one operation initiated by an instruction executed from an executable file and comprising a call to a routine, termed original, external to said executable file,
According to one embodiment of the method, enabling a flexible and only slightly intrusive implementation, the execution of said instruction then produces a call to a routine termed modified, instead of said original routine, this modified routine producing or initiating the implementation of the method.
The modified routine may in particular comprise at least one instruction executed within the system software and which makes a call to at least one logging or replay software agent, ensuring the implementation of the method and being executed in the computer's user memory space of the managed process, or target process.
It is thus possible to produce an implementation of this management essentially within the user space, for example in a way so as to minimise interventions in the system software and to limit risks of errors and of heterogeneity between different computers within the network.
More particularly, the modified routine comprises a test instruction verifying whether the instruction which has called it is executed in the context of a logging or replay, this test thus affecting the call to a management agent of the logging type or of the replay type respectively.
The implementation of the management according to different uses may therefore be produced with improved flexibility by the same agents, for example by a single intermediate application and/or a single kernel module, either for logging or replay.
The prediction function can use different types of software processing, in particular based on the state of the process, its application, or resources which they alone can modify.
Therefore, in one particular embodiment, the method according to the invention uses at least one prediction function comprising a deterministic software processing which depends on the characteristics of one or more external events of at least one type, which occurred previously during the execution of the managed process.
According to one particular feature, the method according to the invention logs at least one first operation by using a first prediction function and a second operation by using a second prediction function.
By using several different prediction functions, it is then possible to adapt the logging method to the circumstances in order to optimise efficiency, with using for the replay of each operation the same prediction function used as for the logging of the corresponding operation.
According to one variation, the method uses a prediction function chosen by deterministic selection or combination within a plurality of given functions, in particular dependent on the type of operation managed, for example according to its nature or according to the resource which is involved. During the replay, this choice of prediction function could thus be reproduced identically or corresponding to the choice made during logging.
According to another variation, the choice of prediction function used depends on the change in a variable measured or calculated during the execution of the logged process, for example a change in the success rate of the prediction function during logging could determine a change of this function or of some of its parameters.
According to the invention, the log file can then comprise data representing at least one choice of prediction function or of parameters affecting such a choice.
Storage of the character and/or the moment of this change during the logging will then enable the production of the same change at the same point of execution during the replay.
In connection with an improvement in the simulation or replay functions, another objective of the invention is to improve the management of the functioning of an application or of at least one of these processes.
The method according to the invention thus produces an functioning management of at least one application process termed logged and comprises the following steps:
In particular, the state of the logged process at the restart point may be captured and stored in the form of restart point data, which is used to restore the restart process to the restart point state before applying the replay step to it.
The method according to the invention may in particular be used to produce a replay of a succession of events comprising external events as well as internal events. The invention thus proposes to carry out the replay of external events by injection or simulation of each of these external events into the restart process. The internal events which follow each external event in the execution of the replay are thus executed by the restart process at its own initiative, in response to the occurrence and/or processing of the external event which immediately preceded them. The restart process thus carries out the replay of at least one logged sequence of internal events in response to the occurrence of an external event initiated by an external replay agent.
Within the functioning management functions, one objective of the invention is also to improve the debugging or analysis or reproduction tools of at least one part of the functioning of an application, or of at least one of the processes from this application.
The functioning management can therefore carry out a monitoring of the execution of an application, termed traced application, for example in connection with the debugging of this application. This monitoring is then applied to at least one process of the traced application and comprises the following steps:
Within the functioning management functions, one objective of the invention is also to improve the tools for reliabilization of the working of an application, or of at least one of these processes.
Such reliabilization may be obtained in particular through the maintenance of improved continuity of operation of the application from the point of view of the services which it provides to its clients. On a failure, this continuity may be total, i.e. the clients do not have to restart the least operation to obtain the same service. Such continuity may also be partial, i.e. by reducing as far as possible the number and/or complexity of the operations which clients will have to repeat or to do in extra to obtain the same service or a part of this service.
Within this objective, the invention proposes to implement functioning management to reliabilize the functioning of a first application, termed reliabilized application, executed in at least one primary node, termed operational node, of a communicating multi-computer architecture, termed cluster.
This reliabilization comprises a restoration, into a state corresponding to the state of the reliabilized application at a restart point, of a second application termed standby, within a second cluster node, termed standby node.
Depending on the embodiments or depending on the situations, the restoration may be done in the absence of any failure and as a preventive measure, or be performed from data stored previously, after detecting a failure.
This reliabilization also comprises the following steps:
The invention also proposes a system comprising a computer implementing such a management method, or a network of computers working together and including at least one such computer.
More particularly, the invention proposes such a network using a middleware type application implementing a method according to the invention to manage the functioning of at least one application executed within said network.
The invention is applicable in particular in a “middleware” type environment, for example producing a network and/or applications management system distributed in one or more networks.
Other features and advantages of the invention will become apparent from the detailed description of a method of embodiment, which is in no way limitative, and the appended drawings in which:
a and 4b illustrate two versions of the operation of a mechanism for the aggregated transmission of logging data from a sequence of internal events;
a and 8b illustrate the operation of logging and replaying internal events so as to process only non-deterministic events;
c and 8d are diagrams illustrating the optimization of internal logging by heuristic compression and decompression respectively;
Within a cluster, an AOP master application, for example a transactional application, provides a certain number of services to users or clients, in particular by input and output of data in different forms. Within the cluster, this application can be single or multi-task (multi-process or multi-threads) and uses a certain number of resources. In particular, these resources can be data, for example in the form of spaces of work memory, shared memory, or data files, or can be state indicators, for example, in the form of semaphores or mutex.
The master application is executed on one or more computers forming a node, termed the operating node OP or the primary node. A functioning management application, termed the intermediate application INT, is executed in one or more parts, in one or more cluster nodes.
According to the embodiments, this intermediate application can deal with different aspects of the master application functioning within the cluster. Such an intermediate application INT can in particular work in parallel with an intermediate cluster management software of the “middleware” type, be integrated with such a middleware, or be itself a form of middleware.
Through the functionalities described here, the intermediate application INT can be used in particular to produce a replication of all or part of a master application in the cluster. The replication of a master application can provide another application which will then be termed replay application.
The functionalities described here, in particular in connection with such replication, also allows the implementation of reliability features for the master application, or tracking or studying this application to implement “debugging”, adjustment or development tasks. A use for reliability implementing will include, for example, the restart application as a backup or replacement application. A use in tracking or debugging will comprise, for example, a logging JOP and/or a replay RSB of events, as described hereafter, according to a slowed or controlled rhythm, of logged events.
The embodiments applied to reliability features are therefore described just here as non-limitative examples.
At different points when running the AOP master application to be made reliable, termed restart points or “checkpoints”, regularly or on event, the intermediate application INT creates or updates at least one restart application ASB executed on a node termed secondary, or “stand by” SB.
This restart application is created or updated, for example by a method of replication by capturing and restoring the application, termed restart method. Said method of replication comprises capture operations CAP of the state of the master application, followed by restoration operations RES of this state, i.e. the state of its processes and of all or part of the resources which it uses.
During such a capture operation CAP, the state of the AOP master application is backed up in the form of data forming a checkpoint state EPR.
Some of the master application's resources, in particular data files representing a large volume on means of storage such as hard disks, can be updated on-the-flow into several copies on several different storage media, constituting restart data files on mirror disks or shared disks. In this case, the data forming a checkpoint state can comprise information constituting references to these restart data files.
When a checkpointing or a replication is based on a capture state including all the execution environment and the master application resources, either directly or by references to replay data files, said checkpoint or said replication can be termed holistic.
From the data of a checkpoint state EPR, the intermediate application INT can implement a restoration RES, by creating or updating a restart application ASB. Said restoration can be implemented regularly or on initiating event, for example at the request of an administrator or a mechanism for managing the cluster workload. This restoration can also be implemented after a failure of the operational node, detected by means of detection, the restart application then being capable of being used as a permanent or non-permanent backup application.
If necessary, the intermediate application organizes a switching of all or part of the master application's services to one or more restart applications. In order that this switching is made transparently for the clients, the intermediate application can use a method of interposition through a “metaprocess” managing virtual network addresses, and implementing a migration of clients' connections from the master application to these restart applications. The intermediate application can also use a method of interposition through a “metaprocess” managing virtual process identifications (virtual PID), allowing the communications for these restart or clone processes to be restored identically to those of their original processes.
These techniques can, for example, be those described in patent FR 2 843 210.
A restoration followed by a partial or total switching can also be implemented besides any fault, for example to distribute the workload of the master application or allow maintenance of certain elements of the operational node or network.
In order that this failure and/or switching are as transparent as possible from the point of view of the clients, the intermediate application records all or part of the events affecting the master application several checkpoints, and backs them up in the form of one or several “logs”.
On completion of a restoration from a checkpoint state, the restart application is in the state of the master application when establishing said checkpoint. Starting from this state, the intermediate application uses the logs that were backed-up since said checkpoint, to cause the re-execution or replay, by the restart application, of the events which have taken place in the master application since this checkpoint. This re-execution, or replay, is implemented so that the restart application reaches the state where the master application was after the last backed-up event, for example, to a state immediately preceding a failure.
The intermediate application can also implement a virtualization of some resources as regards the restart application, for example when these resources have changed since the restored checkpoint, in order to allow it to run its re-execution without interfering with the actual state of these resources, as long as it has not returned to a state corresponding to said actual state.
Among the events to be logged on the operational node and to be replayed on the secondary node, there is a distinction between events termed external and events termed internal.
External events are defined as outside the application in question, in the occurrence the master application. Therefore, external events are defined as those which are initiated in the application by actions or information coming from outside this application, i.e. in particular coming from hardware or software elements which it does not control. These external events can take the form of data or signal inputs, for example hardware interface inputs, such as keyboard or mouse, or data arriving via the network and coming from the outside world, such as a client in the case of a client-server application. Most frequently, these external events cannot be deduced or recreated from the environment of the application. These external events are logged by the master application and can be replayed by the restart application.
If the application in question, sometimes termed the target application, incorporates elements executed on a node other than the primary node, the events in said application but outside the primary node can also be processed as external events.
Internal events are defined as internal to the master application or to the node which is executing it, for example in the form of data or signal inputs received by a process of this application and coming from another process which is also part of the application. These internal events can be received directly or through software mechanisms or agents outside the application but part of the node executing it, for example through partner applications or part of the operating system, such as the “Inter Process Communication” (IPC) agent from a system of the Unix type. These internal events can comprise “message passing events”, for example coming from “pipes”, “signal queues” or “message queues”, or interfaces of the “socket” type. These internal events can also comprise “shared memory access” events, for example semaphores or “mutex”.
While an application is running, internal events are particularly numerous, for example compared with external events. Moreover, internal events correspond to rapid execution, or low latency, operations, in particular compared with the time for a logging operation, above all when the latter includes a network transmission or storage on a permanent medium such as a hard disk. For example, a logging operation can represent a duration 10 to 10,000 times greater than that of an internal event.
As illustrated in
An operational node OP, connected to the cluster by a network, comprises a hardware space, supporting a system space, which itself supports a space termed “user space”. The hardware space, which can be defined by reference to one or more of the lowest layers of the OSI model, comprises in particular hardware devices for executing the process, actual memory and processors and communication, such as network cards. Typically, many external events transit via the hardware space, in the form of communications passing via the network.
The system space, which can be defined by reference to one or more of the intermediate layers of the OSI model, includes, in particular, the operating system. This system space comprises various software mechanisms and agents managing communications from applications with the outside world via the hardware space, for example in the form of sockets in a Unix system, or managing communications between several application processes, for example in the form of a “pipe” and IPC in a Unix system.
The user space, which can be defined by reference to one or more of the topmost layers of the OSI model, comprises processes which are part of the various applications executed by the node, such as the master and intermediate applications. Several processes P1, P2, and Pn, which are part of one or more applications, for example the master application, are executed in this user space. These processes exchange information with the outside through one or more “sockets” from the system space, and between themselves through one or more “pipes” from the system space. Some of these processes also concurrently access “shared memory” resources SHM, in a way managed by state resources (not represented).
When setting up a checkpoint, the intermediate application can start up one or more new logs, or record a “checkpoint mark” in the running log(s).
It should be noted that the term “user”, in particular in the case of the “user space” or the internal events log (“user log”) (described later), is here to be taken to mean “system space user”. This means that the user space is accessible to applications using the node and its operating system, even if this user space is not directly accessible to persons or computers communicating with these applications, and which will then be defined as “clients”.
External events are backed up in a log, made of one or more files KL, termed “kernel log” (c.f.
In the case of a system of the Unix type, said “disp” agent comprises in particular a kernel module, loaded into the system space. This kernel module is loaded dynamically into the kernel when the system is booted, or even before launching the application to be managed or made reliable. From the point of view of the functional structure, for example with reference to the OSI layers, this module is inserted under the IP layer, in particular between the IP layer and the “network device” layer depending on hardware space.
This “disp” agent may intercept and store messages, received from the network and addressed to the TCP layer, in message files QOP and QSB, operating by sending or receiving as required.
In step 1, a message coming from a client, addressed to the process P1, is received by the “disp” agent in the system space of the operational node OP, which retains it in a message queue QOP.
In step 2, a logging message, representing the message received, is sent by the “DISPP” agent from the primary node to a secondary node SB, where a DISPS agent receives it in a received message queue QSB.
The operational node OP may in particular communicate with one or more secondary nodes SB over a separate local area network (LAN), by using a network device different from that used to communicate with clients.
Several of these secondary nodes can also subscribe to an address of the “multicast” type according to the RFC 1112 standard, for communicating with the operational node OP. The use of a multicast address, for example defined by the RFC 1112 standard “Host Extensions for IP Multicasting”) as an IP address in the range situated between 224.0.0.0 and 239.255.255.255, thus allows the operational node to send only once messages addressed simultaneously to a number of secondary nodes, without overloading the network with a transmission which would be sent to all the addresses in the network.
Preferably, the logging message sent from an node OP to another node SB should contain all of the packet or packets received at the physical layer level, in their original form. That is to say, it contains all the data addressed to the master application, as well as network data such as Ethernet, IP, and TCP headers.
In step 3, the secondary node SB sends an acknowledgment message to the operational node OP.
In step 4, on the operational node OP and once the corresponding acknowledgment has been received, the message is retrieved from the message queue QOP and sent to the TCP layer.
In a parallel step 4′, the secondary node SB records the message in a log, for example the kernel external events log KL, and takes it out of the received message queue QSB.
In step 5, in the operational node OP, the P1 process of the master application reads the message in the “socket” element and then processes it for carrying on with its operating.
As the master application only considers the incoming message after acknowledgment by the secondary node SB, the invention ensures that an unlogged message cannot be processed by the application. For example, such an unread message can therefore be retrieved by the TCP protocol's retransmission functions.
If a checkpoint mark is to be set in the kernel log, the control process CtlSB in the secondary node records there the data representing said checkpoint mark.
The content of an internal event depends directly on the local environment, i.e. within the node, on the content of the preceding external events, and on questions of scheduling within a processor or managing a number of processors or computers working in parallel within a node. De facto, in most cases, only the order of these events influences the subsequent behaviour of the application.
The intermediate application INT is confined to logging the order of these internal events, without memorising the detail, or the parameters, of each of them. This choice thus allows the volume of data to be stored for the logging JOP of these internal events to be reduced, and thus to minimise the loss of performance occasioned in the operational node and the master application by this logging.
Internal events are backed up in a log, made of one or more files, termed “user log” (see
As illustrated in
Internal events are scrutinized and read in the operational node's OP user space by a logging process PlogOP of the intermediate application. This then sends to the secondary node's logging process PlogSB, via the high speed connection system HSI, the data representing these internal events and/or their order of occurrence. This data is then backed up in one or more files forming the “User log”.
If a checkpoint mark is to be set in the “user log”, the secondary node's control process PlogSB records there the data representing this checkpoint mark.
Preferably, the logging process PlogOP reads the internal events on their “return”, i.e. when its results have already been produced but not yet sent to the master application process which requested its execution.
This reading is done, for example, by intercepting Input/Output system calls, for example an access to a “pipe”, and the responses to operations locking shared memory segments SHM.
This interception may be implemented by inserting recording instructions (“recording probes”) in the content of all or part of the routines provided by the system and called by the application. These recording probes are added in the form of additional instructions, forming for example an epilogue to the end of the original routine code as illustrated in
The internal events log, the “user log”, thus comprises a succession of records each representing an internal event. These events can be logged in a single file, and will then comprise an identification of the resources and/or processes in question. They can also be recorded in several files, for example one file per resource, or per process, or per combination of the two.
For a file corresponding to a given resource, each of these records comprises, in particular, the following fields:
This result will in particular be used to implement a resources virtualization, for example during the replay of the events in a log by a restart or backup application restored in a secondary node. The stored result will then constitute a value to be forced as the result of an I/O operation request made during the replay, or a virtual identification of a process (virtual PID) in case of a task obtaining a “lock”.
In order to limit losses in performance due to sending logging data from the operational node to one or more secondary nodes, it is useful to aggregate the sending of data representing several internal events.
For this, the intermediate application may use a combination of several different methods, for example implemented by the logging process PlogOP, termed primary, of the operational node OP.
It is understood that the internal change of an application is unimportant as regards to the outside world, for example concerning its clients, as long as this operation does not send anything to the outside world. A restart application, restored from a checkpoint and a log, will not cause any interruption of its services to the outside world if said log does not include internal events which have occurred since the last external message sent by the logged master application.
According to a first way, this primary logging process PlogOP sends internal logging data as it occurs, but in asynchronous mode and according to the transmission availability, without blocking the master application functioning, as long as the latter does not send external messages. On the next sending by the master application of an external message, means of detection warn the primary logging process of this, which then blocks or suspends the sending of this external message, and possibly the execution of one or more of the master application's processes. This blockage is then maintained until all the internal logging data has been sent through this asynchronous transmission, or until it has received a receipt for said data.
According to a second way, the primary logging process PlogOP stores in a buffer or “cache” the internal logging data representing several successive internal events, without sending these immediately to the secondary node's logging process PlogSB. It sends these only when their number reaches a set threshold, or when the application must send a message, termed external, to the outside world, for example data or a signal addressed to a client or to an external process. During the next sending by the master application of an external message, means of detection warn the primary logging process of this, which then blocks or suspends the sending of this external message, and possibly the execution of one or more of the master application's processes. This blockage is then maintained until the primary logging process has sent to the secondary node the remaining logging data in the cache, or until it has received a receipt for this data.
In these two ways, the fact of having to send an external message constitutes an outgoing event, which constitutes a type of event that can be termed blocking, i.e. which requires the logging of preceding events to be closed before execution of this event. According to the embodiments, other types of event can be chosen as blocking, most frequently in addition to external outgoing events.
a illustrates the operation of an events log with aggregation of the logging data DJ for a number of internal events EVI before transmission outside the primary node OP.
In a step 1, the logging process PlogOP detects, during the running of a logged process P1, the occurrence of an event EVI.
In a step 2, the logging process PlogOP checks whether the detected event EVI is of a type that must be considered as blocking.
In a step 3, if the event EVI is not of a blocking type, the logging of this event produces a logging data item DJ.
In a step 4, this logging data item DJ is stored in primary node OP in an ordered structure constituting a buffer log JS1Local, before waiting for the detection of the next event.
In a phase 5, if the detected event EVI is of the blocking type, the logging process PlogOP implements a phase closing the running sequence of internal events previously logged in the buffer log JS1Local.
This phase 5 includes a step 6, where the execution of the logged process P1 is suspended pending satisfactory execution of the closure phase 5.
This phase 5 also comprises a step 7, where the primary node's logging process PlogOP sends the contents of the buffer log JS1Local to the secondary node's logging process PlogSB, which stores it in the log JSem1 concerning the detected event EVI, followed by the preceding data. The primary logging process PlogOP then continues the direct logging of the detected event EVI, with restarting a buffer sequence if this event is also an internal event.
In a variant illustrated in
In this variant, the step 1 of detecting an event is followed by a step b1. In this step b1, if the detected event EVI is of a type considered as initiating, the primary logging process PlogOP checks whether a current sequence SEQC for logging in buffer memory is in progress, and initializes one if this is not the case.
In a subsequent step b2, it tests whether such a current sequence SEQC for logging in buffer memory is in progress for the detected event EVI.
In a step b3, if no current buffer sequence SEQC is active for this EVI event, its result is logged as a logging data item DJ.
In a step b4, said logging data item DJ is sent to the secondary logging process PlogSB, which stores it following those preceding in the log file JSem1 concerning the detected EVI event, following the preceding data. The primary logging process PlogOP then waits for the detection of a new event.
Following step b2, if a current sequence is active for the detected event EVI, the logging of this event continues as in
When the intermediate application wants to switch all or some of the services from the master application to a restart application, it begins by restoring this restart application in a secondary node out of a checkpoint state, and then implements a replay of the events logged since this latter checkpoint.
In particular, for a master application which is event driven, i.e. on initiating events (external), for example a transactional application, the restoration replay is implemented differently for external and internal events.
Such an functioning means, for the application, that it comprises at least one process which can remain to wait to receive an external event, and at this point react by implementing operations comprising internal events
The replay thus comprises an active supply to the application of logged external events, and a passive response providing logged answers in response to internal events, which are created by restart application itself during the replay.
Said kernel log KL is used by a replay process PRE, belonging to the intermediate application and being executed in the user space of the secondary node SB, for re-injecting previously logged TCP messages into a process PB1 of the restart application.
To implement this re-injection, the intermediate application INT comprises or uses a software mechanism or agent which is interposed in the TCP message receipt layers, for example in the form of a software mechanism or agent “ipfilter”, comprising a functional kernel module between the IP and TCP layers. The secondary node also includes a local loop function for the BL network, the access to which is “mapped” by an interface into the system, in order to be accessible to the user space processes. This loop BL can include, in particular, a physical device within the hardware space, which allows it to re-inject the data at the bottom of the IP layer, as opposed to virtual loop-back interfaces which are software implemented in an operating system, for example Unix.
In a step 1, the replay process PRE reads a message logged in the files of the “kernel log” KL.
In a step 2, the replay process PRE injects this message into the of the network local loop BL.
In a step 3, this message is received by the IP layer, which sends it, via the intermediary of the “ipfilter” agent, to the TCP layer for processing.
In a step 4, if the TCP layer sends a receipt to the network, the latter will be filtered or blocked by the “ipfilter” agent.
In a step 5, after having sent the message to the TCP layer, after having received its receipt if any, the “ipfilter” agent signals to the replay process PRE that the message has indeed been received or processed by the TCP layer.
In a step 6, the restart application process PB1 receives the message from the TCP layer, and implements asynchronous reading of the packet(s) contained in it.
Throughout the replay, the “ipfilter” agent isolates the restart application from the network, at the same time preventing all external messages from arriving up to the TCP layer and, at the same time, preventing all messages sent by the restart application from arriving at the IP layer, transparently as regards to this application.
Within the replay application, to implement the replay of internal events occurring between two replayed external events, the intermediate application allows the restart application to run by itself, at the same time virtualizing for it the concerned resources, thus implementing a passive replay. A replay process PRI then detects each operation constituting an internal event as regards to a given resource, and thus forces said resource to adopt the behaviour which has been logged, hence sending to the replay application the result stored for said event during this logging.
During a restoration on a secondary node SB, these two processes PB1, PB2 are implementing a replay based on files constituting a user log. During their replay, the execution of the restart application causes each of these processes to place a call to a single semaphore SEM1, to which corresponds a log file JSEM1 included in the internal events log—the “user log”.
The detection of these access operations and presetting their response are implemented by adding additional instructions in the content of all or part of the routines provided by the systems and called up by the application, using a technique of dynamic interposition by “metaprocess”. Such a technique can be, for example, that described in patent FR 2 843 809. In particular, these instructions can be aggregated before the code implementing the functions from the original routine and thus form a prologue, or be aggregated after this code and form an epilogue.
During execution of the application's executable files, a process P executes a line of code calling up the routine R, for example the routine “sem_wait” from the “POSIX.4” standard, which requests the positioning of a given semaphore for gaining an mutual exclusion access to a given area in shared memory. In the case of a multi-thread application, this may involve an instruction—“pthread_mutex_lock” from the “POSIX threads” standard—which fulfils a similar rôle.
An interposition agent META, loaded into the system when it is booted or before the application's executable files, intercepts the call to the system's original routine R, and forwards it to the amended routine RM. This amended routine comprises instructions implementing or calling up the original routine R—“sem_wait”—preceded by instructions implementing a prologue and followed by instructions implementing an epilogue.
These supplementary instructions can comprise, in particular, an algorithm from the following types:
For the prologue:
if (replay) check(Jsem1)
For the epilogue:
if (replay) end_check(Jsem1)
else record(result, Jsem1)
The instructions “if(replay)” test a condition indicating whether the application is in the process of implementing a replay or not.
In the opposite case (“else”), this means that the application is being executed normally and must therefore be treated as a master application. The epilogue then executes a function—“record(result, Jsem1)”—which is a recording probe as cited previously and participates in the logging of an internal event, at the same time storing the result—“result”—in the log—“Jsem1”.
When the “sem_wait” routine is called up by the restart application during a replay using the “Jsem1” log, the prologue is executed before implementing the system's original “sem_wait” routine.
As the two processes PB1 and PB2 are executed in parallel, their respective calls to the SEM1 semaphore using the “sem_wait” function are not necessarily made in the order stored in the semaphore's log JSEM1.
When the “id2” identifier process PB2 calls up the SEM1 semaphore during the replay, the prologue then executes the instruction “check(Jsem1)” in the name of said same process PB2, in a step 21. This function—“check(Jsem1)”—reads in the JSEM1 log the line corresponding to the current value of the sequence number OPSEM1, that is line “#1: id1”.
This “check” function compares the value PIDlog read, or “id1”, with the identifier of the calling PB2 process, or “id2”. If it notes that these values differ, this “check” function suspends execution of the calling PB2 process, for example by re-executing this same step 21 of comparing in a continuous loop.
Subsequently, when the PB1 process identifier “id1” also calls up the SEM1 semaphore during the replay, the prologue also executes the “check(Jsem1)” instruction, but this time in the name of the new PB1 calling process, in a step 11. If it notes that said PB1 calling process is indeed the one which identifier “id1” is stored in the log at the line corresponding to the current number in the active sequence, i.e. the value “#1”, the “check” function authorises the continued execution of the PB1 calling process.
In a step 12, the amended routine RM then implements the functions of the original routine R, i.e. the “sem_wait” instruction, which then assigns to it the SEM1 semaphore and returns the value “id1” of the PB1 calling process.
In a step 13, the epilogue then executes the “end_check(Jsem1)” instruction in the name of the PB1 calling process. Said “end_check” function then closes the PB1 process' “sem_wait” call and unblocks the execution of the PB2 process which has been on hold. This operation can comprise specifically an incrementing of the sequence number OPSEM1 of this SEM1 semaphore, making it move to the next value “#2”.
In this case, when the “check” function called up by the PB2 process is executed again in a step 22, it reads the next line of the JSEM1 log “#2: id2” and lets its PB2 calling process continue its execution of the amended routine RM.
In a step 23, the amended routine RM then implements the functions of the original routine R, or the “sem_wait” instruction, which then assigns to it the SEM1 semaphore and returns the value “id2” of the PB2 calling process.
In a step 24, the epilogue then executes the “end_check(Jsem1)” instruction in the name of the PB2 calling process, again incrementing the SEM1 semaphore sequence and making it available to continue the replay.
Regardless of the order in which the various replayed processes request assignment of the SEM1 semaphore, it is clear that they can only be obtained in the exact order stored in its JSEM1 log, and therefore in the same order as during the running of the master application which generated this logging.
As these additional instructions are added by a META agent outside the master application and added to the operating system without any change to the latter, it is clear that these logging and replay operations are implemented transparently and non-intrusively, for the master application and without changing pre-existing elements of the system.
Given the large number of internal events, it is useful to optimize the function of their logging and/or replay, in particular to avoid any degradation of performance which would largely detract from the benefits obtained from the aforementioned characteristics.
Among the types of internal event which occur between two external events, most can be classified as deterministic, i.e. incorporating only operations whose results depend exactly on the state of the application before these operations.
On the other hand, in particular during multi-task operations or those distributed over several nodes, some internal events are of a non-deterministic type, as they comprise operations which can provide a result dependent on factors outside the application or the primary node.
By logging or replaying only internal events of non-deterministic types, it is therefore possible to limit overload of the operational node, and thus degraded performance caused by the use of the intermediate application to make reliable or manage the master application.
As illustrated in 8a and 8b, the logging and replay can be accelerated, in particular by only logging the result and only presetting results at replay for internal events where the behaviour is not deterministic.
For all events, and in particular internal events (EVI), a META interposition mechanism (
Whether this involves logging or replay, managing this event EVI then comprises a reactive step to the occurrence of this event, followed by an additional management step GC (
a illustrates the operation of logging an internal event. While the P1 process is executed through a logged (JOP
In a step 1, the amended routine RM corresponding to the event to be logged EVI notifies or initiates the logging process PlogOP, which thus detects the occurrence of this event EVI.
In a step 2, the amended routine RM corresponding to the event EVI implements on the SEM1 semaphore the operation stipulated in the original routine R and receives or calculates a result datum DR addressed to the logged process P1.
In a step 3, the logging process PlogOP increments a sequence number SQ, for example assigned to the SEM1 semaphore, corresponding to the position of the detected event EVI in a P1 logging sequence.
In a step 4, said process PlogOP carries out a test to establish whether the internal event EVI detected is deterministic or not. This test can apply, for example, to a parameter received from the amended routine RM during its call, or to the presence of a result datum DR sent with this call, or to an instruction or event identification previously stored in the primary OP node.
In a step 5, if the detected event EVI is non-deterministic, the PlogOP process sends the result datum DR to the secondary node's PlogSB logging process. This stores the result datum DR and the sequence number SQ corresponding to the event EVI, such that they are associated in a log file JSem1 corresponding to the semaphore SEM1, following the result of preceding logged events. Depending on the logging conditions, data stored in the JSem1 log can also be stored directly in a log file on a permanent medium in the primary node by the PlogOP logging process.
On completion of a sequence of internal events for the logged process P1, the JSem1 log contains an ordered set of all the result data sent by the SEM1 semaphore to said P1 process, associated to the sequence numbers for the events which involve them.
b illustrates the replay operation for this internal event EVI, in a restart process PB1, during a passive replay phase RSB (
In a step 1, the amended routine RM corresponding to the event to be logged EVI notifies or initiates the replay process PRI, which thus detects and identifies the occurrence of this event.
In a step 2, the amended routine RM corresponding to the event EVI implements on the SEM1 semaphore the operation stipulated in the original routine R and receives or calculates a result datum corresponding to the actual replay result RRJ. The amended routine RM then suspends execution of the restart process PB1 and waits for a signal from replay process PRI to send this result RRJ to the restart process PB1.
In a step 3, the replay process PRI reads in the JSem1 log the next unused value RLi for the replay, with the sequence number SQi which is associated to it.
In a step 4, the process for incrementing a sequence number SQ, for example assigned to the SEM1 semaphore, corresponding to the position of the detected event EVI in a PB1 replay sequence PB1.
In a step 5, the replay process PRI carries out a test on the current sequence number SQ and the read sequence number SQi in the log, to establish whether the ongoing replay event EVI corresponds to a logged event.
In a presetting step 7, if these events correspond, the replay process PRI sends the read result RLi in the log to the amended routine RM, which stores it instead of the result RRJ from the original operation R. The amended routine RM then returns this result RLi to the restart process PB1 and lets it continue its execution.
Optionally, the presetting step 7 is preceded by a step 6, in which the replay process PRI receives from the amended routine RM the actual replay result RRJ and compares it with the read result RLi corresponding to the result of the same event during logging. If said two results RRJ and RLi correspond, the process releases directly the amended routine, which returns its result to the restart process PB1 and lets it continue its execution.
Hence it is clear that non-deterministic events can be recorded and replayed faithfully and exactly, ensuring for the restart process PB1a replay run which will be faithful to that of the target process P1 during logging.
As only certain events are logged or replayed, and because supplementary internal operations to implement the invention are much faster then a storage or transmission for logging, overhead due to the operation of the intermediate application INT is reduced.
Optionally, when an original routine R is envisaged only to record events which are deterministic, an amended routine RM which corresponds to it may omit to provide any call to a logging or replay process. Likewise, when an original routine R is envisaged only to implement non-deterministic events, its amended routine RM can comprise a systematic call to a logging or replay process. During logging, step 4 (
In the case where a given type of internal event can be deterministic or not, depending on the type of application or the conditions of its execution, the amended routine RM can also include in its prologue and/or its epilogue instructions evaluating this type of application or these conditions of execution.
The use of a sequence number SQ can also be optional. In this case, the logging process PlogOP (
Moreover, a heuristic, or predictive, method of optimization allows for not logging systematically all internal non-deterministic events. This method can be implemented alone or in combination with other methods of optimization.
Owing to the cost in terms of time for logging and replay operations, in particular as regards to operations internal to a node, it can indeed be useful to implement a certain number of additional internal operations if this allows the number of logging operations to be decreased.
This heuristic optimization technique comprises the implementation, by the intermediate application, of a heuristic compression operating by predicting results and applying over all or part of the internal events detected during the operation of the master application.
During logging in the operational node, this heuristic compression can for example be implemented by the internal logging process PlogOP.
c illustrates the function of the logging of a non-deterministic event, with use of this heuristic compression CH.
While the P1 process is executed through a JOP logged run, executing an instruction implements an internal event EVInD of a non-deterministic type applying to a shared resource, such as a SEM1 semaphore.
In a step 1, the amended routine RMnD corresponding to the event to be logged EVInD notifies or initiates the logging process PlogOP, which thus detects the occurrence of said event EVInD.
In a step 2, the amended routine RMnD corresponding to the event EVInD implements on the SEM1 semaphore the operation envisaged in the original routine RnD and receives or calculates a result datum DR addressed to the logged process P1.
In a step 3, the process PlogOP increments the logging sequence number SQ corresponding to the SEM1 resource involved with detecting the event EVInD.
Advantageously, said sequence number SQ is stored in working memory in the primary node OP. Therefore, its management represents a very low overhead compared with sending a result datum to a secondary node or compared with its storage in a log file on a permanent medium.
This incrementing of the sequence number SQ associated to the SEM1 semaphore and its log JSEM1 thus allows the recording of the passing of a non-deterministic event EVInD correctly predicted by the prediction function FH, while avoiding the overhead which systematic storage of the result datum DR would represent
In a step 4, the process PlogOP implements a software operation FH comprising a prediction of the result of this internal event EVInD in the form a predicted result RP. Preferably, this prediction is a deterministic software process made up of one or more deterministic functions based on the state of the logged process P1 or the master application before this event EVInD.
In a step 5, the process PlogOP compares the predicted result RP with the actual result DR output from the running RnD of the detected event EVInD.
In a step 6, if these two results DR and RP are different, the PlogOP process forwards the actual result DR and the corresponding value of the sequence number SQ, to the secondary node process PlogSB, which memorizes them by associating them as the next line in the log file Jsem1 corresponding to the resource SEM1 in question.
During this step, it is possible to envisage a reinitialization of the sequence number SQ for logging the SEM1 resource in question. In this case, the sequence number SQ represents the number of events correctly predicted since the last event whose result has been logged.
On completion of a sequence of internal events for the logged process P1, the JSem1 log contains an ordered set of all the result data sent by the SEM1 semaphore to said P1 process and which have not been predicted correctly by the prediction function FH.
In the case where the logging of internal events has been implemented by using such an heuristic optimization, the intermediate application then implements a heuristic decompression when replaying in a secondary node. This heuristic decompression uses a prediction identical to that used for compression and applies to the same events as during logging with heuristic compression.
d therefore illustrates the replay operation for a non-deterministic event, with use of this heuristic decompression DH, in a passive replay of a restart process PB1, controlled by an internal replay process PRI (c.f.
During the replay of events from the JSem1 log, executing an instruction implements an internal event EVInD of a non-deterministic type applying to the SEM1 semaphore.
In a step 1, the amended routine RMnD corresponding to the event EVInD to be replayed notifies or initiates the replay process PRI, which thus detects and identifies the occurrence of this event EVInD.
In a step 2, the amended routine RMnD corresponding to the event EVInD implements on the SEM1 semaphore the operation envisaged in the original routine RnD and receives or calculates a result datum corresponding to the actual replay result RRJ. The amended routine RMnD then suspends execution of the replay process PB1. It then waits for a signal from the replay process PRI to forward said result RRJ to the restart process P1 and to let it continue its execution.
In a step 3, the process PRI reads and increments the value of a sequence number SQ corresponding to the semaphore SEM1.
In a step 4, the internal replay process PRI compares this sequence number SQ with the next sequence number SQi not yet replayed from those stored in the log file Jsem1 corresponding to this same resource SEM1.
In a step 5, if these sequence numbers SQ and SQi correspond, then the internal replay process PRI reads the stored result RLi in this log for this sequence number SQi, and stores it as a forced result RF to be returned by the detected event EVInD. The internal replay process PRI then stores the fact that the event represented by the line SQi in the log JSem1 has been replayed, and activates the next line SQj of this same log for the processing of the next event detected.
During this step, it is possible to envisage a reinitialization of the sequence number SQ for replaying the SEM1 resource in question.
In a step 6, if these sequence numbers SQ and SQi do not correspond, the internal replay process PRI implements a software operation FH including the same result prediction as that produced during the logging of this internal event, in the form of a predicted result RPJ. The internal replay process PRI then stores the result RPJ of this prediction as a forced result RF to be returned by the detected event EVInD.
In a step 8, the internal replay process PRI forwards the forced result RF to the amended routine RMnD, which imposes it on the restart process PB1 instead of the actual replayed result RRJ returned by the internal event EVInD. The amended routine then lets the restart process PB1 continue its execution.
Optionally, this presetting can be preceded by a test step 7 to compare these two results RRJ and RF, and avoid acting in the restart process PB1 if these results correspond.
It should be noted that the identification of sequencing data SQ used in this method of predictive optimization can be made up of variables different from those described previously (
It is therefore clear that, even without logging the results of all non-deterministic events, the latter can be recorded and replayed faithfully and exactly. In this way, it is therefore possible to optimize these logging and replay operations, while ensuring a running of the replay run of the restart process PB1d which is faithful to that of the target process P1 during logging.
Given the difference in speed between logging operations and simple calculating operations internal to a node, this heuristic optimization technique can be useful even if the prediction function used does not have a very high success rate. If this difference is large, even a prediction success rate of less than 50% can allow useful optimization.
This heuristic optimization technique can also use several different prediction functions; providing the same is used for logging and then replaying a single event or groups of internal events. The choice of prediction function to be used can be made according to the state of the application or its environment, for example starting from a knowledge database or rules. This change can then be stored in the logging data stored by the intermediate application. This heuristic optimization technique can also be used auto-adaptively, by evaluating its success rate during logging and by initiating a change of said function based on the value of this success rate or on its variation.
An example of the prediction function used in this heuristic optimization technique comprises a prediction of the order of occurrence of internal events based on the order of internal events coming from different clients.
During tasks that are approximately concurrent, and in particular when they have similar or common parts and/or have similar execution times, a prediction function consists of predicting that the order of occurrence of the intermediate internal events Ia1, Ib1, Ic1 will be the same as the order of occurrence of the external events preceding them.
While the master application is running, the order of occurrence of the first external events Eat, Eb1, Ec1 on the operational node OP is recorded by the intermediate application, for example in the internal logging process PlogOP. For example, this order of external events comprises the succession of identifiers of the processes associated to these external events, or the sequence of the values “a b c”.
On each detection of a new internal event concerning this resource, the prediction function makes a prediction of the result of this internal event, i.e. the identity of the process which will obtain the lock over this resource, i.e. the one which has just requested it. This predicted result will then be calculated by comparing the identity of the last process having obtained the lock over this resource, with this order of external events.
Thus, the prediction function will make a set of predictions Pe1 to Pe6, each shown by a dotted line and whose result is shown at its right hand end.
Despite these incorrect predictions Pe3 and Pe6, it is clear that this heuristic optimization will have allowed the internal logging process PlogOP to affect only two transmissions L1 and L2 instead of the six that would have occurred in its absence. This saving of four transmissions out of six represents a work time sufficiently larger than in the internal calculations and operations which are needed to implement this optimization technique, and can therefore provide a significant increase in performance, in particular in the operational node.
Further more, for some internal events where a standard implementation by the operating system will produce a non-deterministic behaviour, it is possible to use a technique of optimization by semantics changing. This technique comprises an amendment to the implementation of such events in the node, in order to give them a behaviour which will be deterministic. The intermediate application makes this amendment identically in the operational node and the secondary node(s), which makes the results of these changed internal events predictable. This amendment to the implementation will be made dynamically by an interposition technique through a “metaprocess” which replaces an original event implementing routine R by an amended routine RM implementing an amended behaviour for this event. The technique used to implement this amendment is similar to that described above (c.f.
This amended behaviour is designed so as to comply with the same specifications as the original behaviour and to be fully compatible with it, for example by planning that, from a given situation where the original routine could have sent a number of different results, the amended routine provides only results which could have been provided by the original routine and are therefore envisaged by the master application and the operating system.
This technique of optimization by semantic change enables a reduction of the number of non-deterministic internal events, whose result must be logged in the operational node to be capable of replay during the restoration of a restart application.
An example of the operation and the interaction of the different parties is illustrated graphically in
A processing agent AT, for example in the system software, carries out an operation which will forward a result DR to a process, for example a logged process P1. For many operations or events, in particular internal, said operation is implemented by an operational process TO, which is by its nature deterministic compared with a set of resources RDet, termed determinant.
From the resources accessible to the process P1, some can be termed reproducible resources RRepr from knowledge of the state of this process P1. Said reproducible resources comprise in particular resources whose state depends exclusively on it.
In the operation of the processing agent AT, processing the TO operation can comprise a processing part TD which is deterministic as regards to reproducible resources RRepr of the process P1, for example because it uses only DER data coming from said reproducible resources.
In the case where the operation process TO comprises another part of processing using personal data coming from SEM1 resources not included in the reproducible resources RRepr of the process P1, it is common for the result of this TnD part, and thus all the TO processing, not to be deterministic as regards to the process P1 invoking it.
In such a situation, this semantic change technique may consist of using a management agent AG to amend the processing agent's behaviour or the data which it uses or produces, in order that the operation resulting from this amendment is deterministic compared with the reproducible resources RRepr.
This managing agent can use a functioning modifying treatment TMF to amend the internal operation of the TO operating process.
It can also use the input data DE output from the determinant resources RDet but not reproducible (RRepr) as regards to the process P1, to compensate for variations to the result DR capable of constituting a non-deterministic source for said same process P1. Such a compensation can be carried out by TC1 modifying the input data DE into compensated input data DEC, or by TC2 modifying the result data DR into compensated result data DRC.
This managing agent AG can also choose or regulate the modifications TMF, TC1, TC2 which it made, depending on one or more semantic change parameter PCS, in order to optimize the efficiency of the global processing AT and AG. In order to remain reproducible between a logging JOP and a replay RSB, it is sufficient that the variations to this semantic change parameter PCS are determined only by data from the reproducible resources RRepr, or that its variations are stored in the logs UL, KL during logging and are read and applied in the same way during the replay RSB
This change of behaviour can concern in particular aspects affecting the management of several processes competing for a given resource.
In its standard implementation, the “read” routine, initiated by an application, uses a zone of buffer memory B to read messages in an input channel ICH and forward these to said application. The messages are received in the system in the form of successive data which are stored in a memory zone forming the input channel, as and when they arrive. According to its configuration, the “read” operation can use a buffer of different sizes, but this buffer is used in its entirety for each reading in the input channel.
In this example, the application uses a succession of “read” operations for the buffer B of size “50” to receive three messages M1, M2, M3 which reach it successively through the input channel ICH. These three messages represent data volumes equaling “20”, “30” and “50” respectively. However, the speed at which data arrives in the input channel, on the one hand, and the speed of reading operations, on the other hand, can vary from one another in a way which cannot be predicted at the stage of a logging or replay.
In a first scenario SCA, a first reading RA1 is made as only the data from the first message M1, size “20”, has arrived. The buffer B is not completely filled, and the operation returns a result corresponding to the content “M1” and a data size of “20”. A second reading RA2 is then made after the arrival of the second message M2 alone, which returns a result corresponding to the content “M2” and a data size “30”. A third reading RA3 is then made after the arrival of the third message M3, which returns a result corresponding to the content “M3” and a data size “50”. For example, for the size of data received by the application, this first scenario A thus returns a set of three results equal to “20, 30, and 50”.
In a second scenario SCB, a first reading RB1 is taken as the same first and second messages M1, M2 have already arrived, which returns a result corresponding to the content “M1, M2” and a data size of “50”. A second reading RB2 is then made after the arrival of the third message M3, which returns a result corresponding to the content “M3” and a data size of “50”. For the size of the data received by the application, this first scenario SCA thus returns a set of two results equaling “50, 50”, this for reading the same messages.
These two scenarios therefore return different results “20, 30, 50” for one and “50, 50” for the other. In this, the standard system routine implementing the “read” operation implements a non-deterministic event from the point of view of the application, for logging the master application as well as for replaying a restart application.
For the same situation as in
In this example, the amended routine recognizes the actual length of each of the messages received and reads in the input channel ICH only the data corresponding to a single message, even if the buffer B is not filled and there is still data to read in the input channel ICH. In the case of the logging of the master application, the amended routine recognizes the actual length of the messages M1, M2, M3 using an external event logging mechanism corresponding to the receipt of these same messages, for example the IPlogOP module. In the case of a replay while the restart application is being restored, the amended routine recognizes the actual length of the messages M1, M2, M3 using an external event replay mechanism corresponding to the receipt of these same messages, for example the IPlogSB module.
These two different arrival scenarios SCA, SCB thus give a single behaviour for the reading operation, in the occurrence of a single set of three results equaling “20, 30, 50” for the size of the data received by the application.
Likewise, for other sizes of the buffer B, an original “read” routine producing different sets of results is possible.
Thus, the following results can be obtained for a buffer size of “20”, for example: “20, 20, 20, 20, 20” or “20, 20, 10, 20, 20, 10”.
The following results can be obtained for a buffer size of “100”, for example: “20, 30, 50” or “50, 50” or “20, 80” or “100”.
On the other hand, for each buffer size, a “readM” routine thus amended can give only a single set of results.
Thus, for a buffer size “20”, the set of results obtained will be “20, 20, 10, 20, 20, 10”.
For a buffer size “100”, the set of results obtained will be “20, 30, 50”.
The “readM” routine thus amended therefore implements a deterministic behaviour for the internal event corresponding to such a reading operation.
In this example, three messages M1, M2, M3 with contents equaling “a”, “b” and “c” respectively are received by the node operating system OS, addressed to two different channels ICH1, ICH2.
This example can be applied in particular to the receipt of data in the form of a “stream” by the first channel ICH1, and data in the form of messages or packets of the TCP type by the second channel ICH2. In the operating system OS, two TCP packets, followed by a “stream” packet are then received as three successive messages M1, M2, M3 with content equaling “a”, “b” and “c” respectively.
As it receives these and according to its workload, the operating system OS processes and distributes this data in the channels ICH1, ICH2 corresponding to their type. At a given instant during its execution, the application calls up the “select” routine to initiate a reading operation for the different channels by which it can receive messages.
In its standard implementation, the “select” routine reads the queuing data in the first channel ICH1, followed by that in the second channel ICH2, and forwards this immediately to the application, in the order in which it read them.
Now, the speed at which data arrives in the operating system OS, the speed of its processing by the operating system and therefore its speed of arrival in the input channels, on the one hand, and the speed of execution by the application of a succession of read operations, on the other, can vary from one another in a way which cannot be predicted at the stage of a logging or replay.
In a first scenario SCA, illustrated in
In a second scenario SCB, illustrated in
These two different scenarios SCA, SCB therefore return different results, “c a b” for one and “a b c” for the other. In this, the standard system routine implementing the “select” operation implements a non-deterministic event from the point of view of the application, for logging the master application as well as for replaying a replay application.
For the same situation as in
In this example, the amended routine recognizes the order in which the messages arrive in the operating system OS, and reads the messages in the order in which they arrived. Moreover, to reduce the risks of ambiguity, the amended routine sends only a single file descriptor each time. The amended routine can obtain information on the order in which the messages arrive, for example by examining the content of messages in the input channels ICH1, ICH2, or from logging or replay data.
These two different arrival scenarios SCA, SCB thus give a single behaviour for the multiplexed reading operation, in the event a single set of three results equaling “a b c”.
By thus amending the method of operation of certain routines implementing the behaviour of internal events which were not deterministic in a standard environment to make them deterministic, it is clear that a reduction of the number of non-deterministic events is obtained. When this amendment is applied identically during logging in the master application and during replay in a restart application, the number of events which must be logged to be able to obtain, on completion of the replay, a restart application which is in a state corresponding to that of the master application or which has a satisfactory continuity of operation with said master application, is reduced.
Therefore, it is clear that this technique of optimization by semantic change can improve the performance of the logging and replay operations, and thus the intermediate application.
In fact, according to the routines to which this technique of semantic change is applied, and according to the nature of the amendment made to them, this can result in a minor fall-off in performance in said routine compared with its original behaviour. However, given the slow speed of logging operations, the savings generated in terms of the number of operations to be logged can allow significant improvement in the overall performance of the master application within the framework of the intermediate application.
In this description, it can be seen that the mechanisms of the intermediate application are in the main implemented by processes or modules being executed in the user space of the operational node or of secondary nodes. In particular, this means the logging or replay processes, external or internal, identified here in the intermediate application INT (
In contrast, the mechanisms being executed in the system space comprise above all the interposition modules, or those for adding or amending functions, which are managed from the application modules. This in particular means modules identified here under the references “DISP” (
The fact that the execution and the “life” of the intermediate application takes place in the user space enables to limit the interactions with the operating systems of the different nodes. This feature in particular provides a flexibility in deployment and management, a certain independence towards operating systems and their optional heterogeneity, limits the risks of type or release incompatibility, and can limit interventions in the system space of the nodes which are not involved, or only to a lesser degree, in the deployment of said intermediate application. This independence towards operating systems can also limit development times and costs, by avoiding too much in-depth intervention in pre-existing elements of the system space, and retaining a certain commercial and technical independence towards specifications and changes to these operating systems and towards the policies of the organizations which manage them.
An intermediate application as described above can be implemented in different ways and according to different combinations to provide users or managers of a cluster with a support or management service for other applications. Such a service can in particular be obtained in the form of a network software product of the “middle-ware” type, allowing the management, optimizing or reliability improvement, in a cluster, of one or more applications in their original version (“legacy”), at the same time providing functions of flexibility or additional security or fault tolerance, for example adapted to the nature of the cluster.
The use of such an intermediate application can more particularly take the form of securing services provided by these applications to their clients. Each application will therefore be able to be treated as a master application and be restored in the form of a restart application to replace the master application for its clients as necessary.
The services provided by the applications being executed in all or part of a given node can also be shifted to one or more other nodes dynamically and on demand, by freeing their original node completely. Therefore, it will be possible to implement all the hardware or software interventions wanted on this node, whether these be maintenance, trials, upgrades or replacement.
Such an intermediate application can be used in particular to implement an environment of the “middleware” type, comprising functionalities for distributing the workload (load balancing) between the different nodes, to optimize the use of the different hardware, in particular according to its power, availability, or its geographic situation in the network, for example its remoteness from its clients or the data which is used.
Obviously, the invention is not limited to the examples which have just been described and numerous amendments may be made thereto, without departing from the framework of the invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 05 00610 | Jan 2005 | FR | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/EP2006/050339 | 1/20/2006 | WO | 00 | 7/13/2007 |
| Publishing Document | Publishing Date | Country | Kind |
|---|---|---|---|
| WO2006/077247 | 7/27/2006 | WO | A |
| Number | Name | Date | Kind |
|---|---|---|---|
| 5363503 | Gleeson | Nov 1994 | A |
| 5802272 | Sites et al. | Sep 1998 | A |
| 6182086 | Lomet et al. | Jan 2001 | B1 |
| 6732307 | Edwards | May 2004 | B1 |
| 7321992 | Vellore et al. | Jan 2008 | B1 |
| 20030040887 | Shupps et al. | Feb 2003 | A1 |
| 20030177411 | Dinker et al. | Sep 2003 | A1 |
| 20060095734 | Filippo et al. | May 2006 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20080086730 A1 | Apr 2008 | US |
