Patent Grant
7383317
A portion of the disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document of the patent disclosure, as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever.
The following U.S. patent applications are cross-referenced and incorporated herein by reference:
U.S. patent application Ser. No. 09/975,590 entitled “DATA REPLICATION PROTOCOL” by Dean Bernard Jacobs, Reto Kramer, and Ananthan Bala Srinivasan, filed Oct. 11, 2001; and
U.S. patent application Ser. No. 10/234,597, entitled “EXACTLY ONCE JMS COMMUNICATION” by Dean Bernard Jacobs and Eric Halpern, filed Sep. 4, 2002, issued Sep. 26, 2006 as U.S. Pat. No. 7,113,980.
The present invention is related to technology for distributing objects among servers in a network cluster.
In distributed computer systems, it is often the case that several servers and/or networking nodes need to work together. These servers and nodes have to be coordinated, as there is typically networking information that needs to be shared among the machines in order to allow them to function as a single entity. Typical approaches to machine coordination can be very expensive in terms of resources and efficiency.
In general, some synchronization is required for the nodes to agree, as there may be several messages passing between the nodes. This requirement for synchronization may, however, be undesirable in a clustered networking environment. Many clustered environments simply avoid imposing any such synchronization requirement. There are applications, however, where agreement is necessary.
In one case where agreement is needed, a device can exist to which a cluster may want exclusive access. One such device is a transaction log on a file system. Whenever a transaction is in progress, there are certain objects that need to be saved in a persistent way, such that if a failure occurs those persistently-saved objects can be recovered.
For these objects that need to be saved in one place, there is typically a transaction monitor that runs on each server in that cluster or domain, which then uses a local file system to access the object. Each server can have its own transaction manager such that there is little to no problem with persistence. There is then also no need for coordination, as each server has its own transaction manager.
For example, there can be a cluster including three servers, each server having a transaction manager. One of those servers can experience a failure or other problem causing the server to be unavailable to the cluster. Because the failed server is the only server having access to a particular transaction log, none of the transactions in that particular log can be recovered until the server is again available to the cluster.
Recovery of the log can be difficult or at least inefficient, as a problem with the server can take a significant amount of time to fix. Significant server problems can include such occurrences as the shorting out of a motherboard on the server or a power supply being burnt out.
The present invention includes a system for managing objects, such as can be stored in servers on a network or in a cluster. The system includes a data source, application, or service, such as a file system or Java Message Service component, which can be located inside or outside of a cluster. The system can include several servers in communication with the file system or application, such as through a high-speed network connection.
The system includes a lead server, such as can be agreed upon by the other servers. The lead server can be contained in a hardware cluster or in a software cluster. The system can include an algorithm for selecting a lead server from among the servers, such as an algorithm built into a hardware cluster machine. The lead server in turn will contain a distributed consensus algorithm for selecting a host server, such as a Paxos algorithm. The algorithm used for selecting the lead server can be different from, or the same as, the algorithm used to select the host server.
The host server can contain a copy of the item or object, such as can be stored in local cache. The host server can provide local copy access to any server on the network or in a cluster. The host server can also provide the only access point to an object stored in a file system, or the only access point to an application or service. Any change made to an item cached, hosted, or owned by the host server can also be updated in the file system, application, or service.
If the host server becomes unable to host the object, a new host can be chosen using a distributed consensus algorithm. The new host can then pull the necessary data for the object from the file system or service. The other servers in the cluster can be notified that a new server is hosting the object. The servers can be notified by any appropriate means, such as by point-to-point connections or by multicasting.
Systems in accordance with the present invention can provide solutions to availability issues, such as when a server owning a data object becomes unavailable to a server cluster. One such solution allows for another server in the cluster to take over the ownership of the data object. A problem arises, however, in making the data object accessible to both servers without having to replicate the data object on both.
If a file system, data store, or database (all collectively referred to hereinafter as “file system”) is used by the cluster to persistently store data, and the file system is accessible from more than one server, the second server can automatically take over the task of data object access if the first server owning that object encounters a problem. Alternatively, there can be an algorithm utilized by the cluster or a server in the cluster to instruct a server to take ownership of the item. Another fundamental problem, however, involves getting the cluster to agree on which server now owns the resource or object, or achieving a “consensus” amongst the servers.
A server can “own” or “host” a data object when, for example, it acts as a repository for the object, such as by storing a copy of the data object in local cache and making that copy available to other servers in the cluster, or by being the sole server having direct access to an object in a file system, service, or application, such that all other servers in the cluster must access that object through the hosting server. This ensures that an object exists “exactly once” in the server cluster.
If a hosted object is, for example, to be cached on the hosting server, a copy of the data object can be pulled from a file system to the host server and stored as an object in local cache 304. The other servers on the network or in the appropriate cluster are then notified, such as by the hosting server, that a local copy of the object exists on the hosting server, and that the local copy should be used in processing future network requests 306.
In a Paxos algorithm, one example of a distributed consensus algorithm, a server can be selected to act as a host or lead server by a network server, the network server leading a series of “consensus rounds.” In each of these consensus rounds, a new host or lead server is proposed. Rounds continue until one of the proposed servers is accepted by a majority or quorum of the servers. Any server can propose a host or lead server by initiating a round, although a system can be configured such that a lead server always initiates a round for a host server selection. Rounds for different selections can be carried out at the same time. Therefore, a round selection can be identified by a round number or pair of values, such as a pair with one value referring to the round and one value referring to the server leading the round.
The steps for one such round are as follows, although other steps and/or approaches may be appropriate for certain situations or applications. First, a round can be initiated by a leader sending a “collect” message to other servers in the cluster. A collect message collects information from servers in the cluster regarding previously conducted rounds in which those servers participated. If there have been previous consensus rounds for this particular selection process, the collect message also informs the servers not to commit selections from previous rounds. Once the leader has gathered responses from at least half of the cluster servers, for example, the leader can decide the value to propose for the next round and send this proposal to the cluster servers as a “begin” message. In order for the leader to choose a value to propose in this approach, it is necessary to receive the initial value information from the servers.
Once a server receives a begin message from the leader, it can respond by sending an “accept” message, stating that the server accepts the proposed host/lead server. If the leader receives accept messages from a majority or quorum of servers, the leader sets its output value to the value proposed in the round. If the leader does not receive majority or quorum acceptance (“consensus”) within a specified period of time, the leader can begin a new round. If the leader receives consensus, the leader can notify the cluster or network servers that the servers should commit to the chosen server. This notification can be broadcast to the network servers by any appropriate broadcasting technology, such as through point-to-point connections or multicasting.
The agreement condition of the consensus approach can be guaranteed by proposing selections that utilize information about previous rounds. This information can be required to come from at least a majority of the network servers, so that for any two rounds there is at least one server that participated in both rounds.
The leader can choose a value for the new round by asking each server for the number of the latest round in which the server accepted a value, possibly also asking for the accepted value. Once the leader gets this information from a majority or quorum of the servers, it can choose a value for the new round that is equal to the value of the latest round among the responses. The leader can also choose an initial value if none of the servers were involved in a previous round. If the leader receives a response that the last accepted round is x, for example, and the current round is y, the server can imply that no round between x and y would be accepted, in order to maintain consistency.
A sample interaction between a round leader and a network server involves the following messages:
There can be a file system that is separated from the servers, located either inside or outside of a hardware or software cluster. This file system can persistently store the transaction log, such as by storing the log on a first disk and replicating the log to a second disk within the file system. If the first disk crashes, the file system can hide the crash from the cluster and/or server and get the log information from the second disk. The file system can also choose to replicate the log to a third disk, which can serve as a backup to the second disk.
From the perspective of a server in the cluster, the file system can be a single resource. In one embodiment, the server may only care that a single server owns the file system at any time.
Another example of a system in accordance with the present invention involves caching in a server cluster. It may be desirable in a clustered environment, such as for reasons of network performance, to have a single cache represent a data object to servers in the cluster. Keeping items in a single cache can be advantageous, as servers in the cluster can access the cache without needing to continually return to persistent storage. Being able to pull an item already in memory can greatly increase the efficiency of such a system, as hits to a database or file system can be relatively time intensive.
One problem with a single cache, however, is that it may be necessary to ensure that the object stored in memory is the same as that which is stored on a disk of the file system. One reason for requiring such consistency is to ensure that any operations or calculations done on a cached item produce the correct result. Another reason is that it can be necessary to restore the cache from the file system in the event that the cache crashes or becomes otherwise tainted or unavailable.
There can be at least two primary ways to handle this type of caching in a cluster, although other ways may work at least as well for certain applications. One way is to replicate the cache in multiple places. This approach can be problematic, as any change to an item being cached requires that all servers replicating the cache agree to, or are at least aware of, the change. This can prove to be very expensive in terms of resources and performance.
An alternative approach in accordance with the present invention assigns a particular server to be the owner of a cache in the cluster, and all access to the cache goes through that particular server. Any server in a cluster can host such a cache. Each server can host one, several, or no caches. The caches can be hosted on a single server, or spread out among some or all of the servers in the cluster. The cluster itself can be any appropriate cluster, such as a hardware cluster or a group of servers designated by a software application to be in a given “software” cluster.
It may be possible to think of either example, a transaction log and/or a cache, as a type of object that sits somewhere on a system. It may be desirable to ensure that any such object exists only once in a cluster, and that the object is always available. It may also be desirable to ensure that the object can be recovered on another server if the server hosting the object fails, and that the object will be available to the cluster.
One method 400 for recovery is shown in
Systems and methods in accordance with the present invention can define objects that exist in exactly one place in a cluster, and can ensure that those objects always exist. From a server's perspective, it may not matter whether an object such as a transaction log is mirrored or replicated, such as by a file system. From the server's perspective, there is always one persistent storage accessible by any server in the cluster. The system can periodically check for the existence of an object, or may assign ownership of objects for short periods of time such that an object will be reassigned frequently to ensure existence on some machine on the network or in the cluster.
A hardware cluster can comprise a bank of machines, each machine being capable of running multiple servers. There can also be a file system behind each machine. Servers in a hardware cluster are typically hardwired, such that they are able to more quickly make decisions and deal with server faults within the hardware cluster. Hardware clusters can be limited in size to the physical hardware of the machine containing the servers. Servers in a hardware cluster can be used as servers in a software cluster, and can also comprise network servers, as the individual servers on the machines are available to the network.
The shared file system for one of these machines can be available to all servers in a cluster, such as through a high-speed network. The file system can also be redundant. In one embodiment, this redundancy is implemented through the use of multiple data disks for the file system. In such a redundant implementation, an object can be replicated across multiple disks any time the object is written to the file system. Such a file system, when viewed as a “black box,” can be able to withstand failures of any of the disks and still provide access to a data item from any of the servers in the cluster.
A framework in accordance with the present invention, referred to as an “exactly-once”framework, can be built on the assumption that these objects kept in memory are always backed by a reliable, persistent storage mechanism. For example, there can be an object that represents a transaction log. Whenever a call is made to the object, the corresponding transaction log can be updated. This may include a call that either reads from, or writes to, a database. An object representing that transaction log can be sitting on one of the servers in the cluster, such as the host server. An exactly-once framework can ensure that, as long as at least one of the servers in the cluster is up and running, a server will be able to take over ownership of the log if another server fails.
There may be one object that represents a cache. Every time the cache is updated, the update can also be written back to persistent storage. When one of the servers needs to use a data item, that server can be required to go through this object. If the server hosting the object that represents the cache fails, the object can be resurrected on another server. The resurrected object can pull all the necessary information from persistent storage.
An exactly-once framework can act as a memory buffer for use by the cluster. The framework can provide a single cache representing data in the system that is backed by a reliable, persistent storage. Whenever data is read from the cache, the read can be done without needing to access persistent storage. When an update is written to cache, however, it can be necessary to write back through the persistent storage, such that the system can recover if there is a failure.
One important aspect of an exactly-once framework involves the way in which the approach is abstracted, which can vary depending upon the application and/or implementation. A new type of distributed object is created, such as can be referred to as an “exactly-once object.” An exactly-once object can be, for example, a locally-cached copy of a data item in a file system, or the sole access point to such a data item for servers in a cluster. Underlying techniques for implementing this abstraction can also be important.
Systems of the present invention can utilize any of a number of methods useful for distributed consensus, such as a method using the afore-mentioned Paxos algorithm. Such an algorithm can be selected which provides an efficient way for multiple nodes and/or distributed nodes to agree on one value of an object. The algorithm can be chosen to work even if nodes fail and/or return during an agreement process.
A typical approach to network clustering utilizes reliable broadcasting, where every message is guaranteed to be delivered to its intended recipient, or at least delivered to every intended functioning server. This approach can make it very difficult to parallelize a system, as reliable broadcasting requires a recipient to acknowledge a message before moving on to the next message or recipient. A distributed algorithm utilizing multicasting may reduce the number of guarantees, as multicasting does not guarantee that all servers receive a message. Multicasting does simplify the approach such that the system can engage in parallel processing, however, as a single message can be multicast to all the cluster servers concurrently without waiting for a response from each server. A server that does not receive a multicast message can pull the information from the lead server, or another cluster server or network server, at a later time. As used herein, a network server can refer to any server on the network, whether in a hardware cluster, in a software cluster, or outside of any cluster.
An important aspect of an exactly-once architecture is that consensus difficulties are reduced. In accordance with the present invention, the performance of a distributed consensus implementation can be improved by using multicast messaging with a distributed consensus algorithm. This approach can allow for minimizing the message exchange and/or network traffic required for all the servers to agree.
When multicasting, one of several approaches can be taken. In a first approach, which may be referred to as “one-phase distribution,” a lead server can multicast a message to all other servers on the network, such as may be used in a round of a Paxos algorithm, or used to state that a new host has been selected for an object. In this approach, the lead server only needs to send one message, which can be passed to any server available on the network. If a server is temporarily off the network, the server can request the identification of the new host after coming back onto the network.
Using another multicast approach, which may be referred to as a “two-phase distribution,” the lead server can pre-select a host server using an appropriate algorithm. Before assigning an object to that host, however, the lead server can contact every other server in the cluster to determine whether the servers agree with the choice of the new host server. The lead server can contact each server by a point-to-point connection, or can send out a multicast request and then wait for each server to respond. If the servers do not agree on the selection of the host, the lead server can pre-select a new host using the algorithm. The lead server would then send out another multicast request with the identity of the newly pre-selected host in another round.
If every server agrees to the pre-selected host, the lead server can assign the object to the host server. The lead server can then multicast a commit message, informing the servers that the new change has taken effect and the servers should update their information accordingly.
An exactly-once framework can also utilize a “leasing” mechanism. In using such a mechanism, an algorithm can be used to get the cluster servers to agree on a lead server, such as by using distributed consensus. Once selected, that lead server can be responsible for assigning exactly-once objects to various servers in the cluster. The system can be set up such that the cluster servers will always agree on a new leader if an existing lead server fails.
While the lead server is active, the lead server can be aware of all the exactly-once objects that need to exist in the system. The lead server can decide which server should host each object, and can then “lease” that object to the selected server. When an object is leased to a server, that server can own or host the object for a certain period of time, such as for the duration of a lease period. The lead server can be configured to periodically renew these leases. This approach can provide a way to ensure that a server will not get its lease renewed if it fails or becomes disconnected in any way, or is otherwise not operating properly within the cluster.
Much of the problem with distributed systems under failure is that it is difficult to tell the difference between a server that has failed and one that is simply not responding. Any server that has been somehow cut off the network can no longer host an object. That server will still know, even though it is not available to the cluster, that it can drop its hosting of any object after the lease period. As the server is not available to the cluster, it will not get its lease renewed.
The lead server also knows that, if the lead server is unable to reach the host server within a certain amount of time, the host server will relinquish its ownership of the object. The lease period can be for any appropriate time, such as for a matter of seconds. The lease period can be the same for all objects in the cluster, or can vary between objects.
A system using an exactly-once architecture can also be tightened down. Operating systems often provide special machinery that is built closer to the hardware and can offer more control. One problem with this approach, however, is that it can be limited by the hardware available. For example, a hardware cluster of servers can have on the order of 16 servers. Because these systems require some tight hardware coupling, there can be limitations on the number of servers that can be included in the cluster.
An exactly-once framework, on the other hand, may be able to handle clusters much larger than these proprietary hardware clusters can handle. A framework can allow for some leveraging of the qualities of service that are available from one of the proprietary clusters, thereby allowing for a larger cluster. Differing qualities of service may include, for example, whether messages are sent by a reliable protocol, such as by point-to-point connections, or are sent by a less reliable but more resource-friendly protocol, such as multicasting. An advantage to using an exactly-once framework is the ability to balance scalability with fault tolerance, such that a user can adapt the system to the needs of a particular application.
Prior art systems such as hardware cluster machines can attempt high availability solutions by having (what appears to the cluster to be) a single machine backed up by a second machine. If the first machine goes down, there is a “buddy” that takes over, and any software that was running on the first machine is brought up on the second machine.
An exactly-once framework in accordance with the present invention can assign the lead server to a server in one of these hardware clusters, such that dealing with leader failure can become faster than dealing with it in a software cluster. This lead server can, however, dole out leases to servers whether those servers are in the hardware cluster or the software cluster. This arrangement may provide for faster lead server recovery, while allowing for a software cluster that is larger than, but still includes, the hardware cluster.
One such system 200 is shown in
One method 500 for using such a system is shown in
An exactly-once framework can be used, for example, to handle transaction logs or caching. Such a framework can also be used, for example, to define an administration server as an exactly-once object and lease the administration server such that the administration server never goes down.
The system of
Portable storage medium drive 714 operates in conjunction with a portable non-volatile storage medium, such as a floppy disk, to input and output data and code to and from the computer system of
User input devices 712 provide a portion of a user interface. User input devices 712 may include an alpha-numeric keypad for inputting alpha-numeric and other information, or a pointing device, such as a mouse, a trackball, stylus, or cursor direction keys. In order to display textual and graphical information, the computer system of
The components contained in the computer system of
The foregoing description of preferred embodiments of the present invention has been provided for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Obviously, many modifications and variations will be apparent to the practitioner skilled in the art. The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications that are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the following claims and their equivalence.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4714996 | Gladney | Dec 1987 | A |
| 5163148 | Walls | Nov 1992 | A |
| 5613060 | Britton | Mar 1997 | A |
| 5765171 | Gehani | Jun 1998 | A |
| 5796934 | Bhanot et al. | Aug 1998 | A |
| 5805798 | Kearns | Sep 1998 | A |
| 5909689 | Van Ryzin | Jun 1999 | A |
| 6018805 | Ma | Jan 2000 | A |
| 6112257 | Mason | Aug 2000 | A |
| 6119143 | Dias | Sep 2000 | A |
| 6122629 | Walker | Sep 2000 | A |
| 6134673 | Chrabaszcz | Oct 2000 | A |
| 6212521 | Minami | Apr 2001 | B1 |
| 6263368 | Martin | Jul 2001 | B1 |
| 6327622 | Jindal | Dec 2001 | B1 |
| 6389462 | Cohen | May 2002 | B1 |
| 6425005 | Dugan | Jul 2002 | B1 |
| 6438705 | Chao | Aug 2002 | B1 |
| 6502125 | Kenner | Dec 2002 | B1 |
| 6505200 | Ims et al. | Jan 2003 | B1 |
| 6505241 | Pitts | Jan 2003 | B2 |
| 6523130 | Hickman | Feb 2003 | B1 |
| 6526521 | Lim | Feb 2003 | B1 |
| 6539381 | Prasad et al. | Mar 2003 | B1 |
| 6578160 | MacHardy, Jr. et al. | Jun 2003 | B1 |
| 6606643 | Emens | Aug 2003 | B1 |
| 6721777 | Sharma | Apr 2004 | B1 |
| 6757708 | Craig et al. | Jun 2004 | B1 |
| 6775703 | Burns | Aug 2004 | B1 |
| 6779017 | Lamberton | Aug 2004 | B1 |
| 6826601 | Jacobs et al. | Nov 2004 | B2 |
| 6832238 | Sharma | Dec 2004 | B1 |
| 6877111 | Sharma et al. | Apr 2005 | B2 |
| 6944785 | Gadir | Sep 2005 | B2 |
| 6963857 | Johnson | Nov 2005 | B1 |
| 7100195 | Underwood | Aug 2006 | B1 |
| 7111061 | Leighton | Sep 2006 | B2 |
| 7113980 | Jacobs | Sep 2006 | B2 |
| 7171692 | DeMello | Jan 2007 | B1 |
| 20010042073 | Saether et al. | Nov 2001 | A1 |
| 20010054062 | Ismael et al. | Dec 2001 | A1 |
| 20020073188 | Rawson | Jun 2002 | A1 |
| 20020073354 | Schroiff et al. | Jun 2002 | A1 |
| 20020147961 | Charters | Oct 2002 | A1 |
| 20020161839 | Colasurdo | Oct 2002 | A1 |
| 20020188591 | Santosuosso | Dec 2002 | A1 |
| 20030041135 | Keyes et al. | Feb 2003 | A1 |
| 20030060214 | Hendrey et al. | Mar 2003 | A1 |
| 20030065686 | Callahan et al. | Apr 2003 | A1 |
| 20030088713 | Mandal et al. | May 2003 | A1 |
| 20030233433 | Halpern | Dec 2003 | A1 |
| 20030236923 | Jeyaraman | Dec 2003 | A1 |
| 20040059735 | Gold et al. | Mar 2004 | A1 |
| 20040139125 | Strassburg et al. | Jul 2004 | A1 |
| 20040153558 | Gunduc et al. | Aug 2004 | A1 |
| 20040230747 | Ims et al. | Nov 2004 | A1 |
| 20060168118 | Godlin et al. | Jul 2006 | A1 |
| 20060212453 | Eshel et al. | Sep 2006 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20040249928 A1 | Dec 2004 | US |
| Number | Date | Country | |
|---|---|---|---|
| 60317718 | Sep 2001 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10234693 | Sep 2002 | US |
| Child | 10887720 | US |
