Patent Application
20050246581
System availability, scalability, and data integrity are fundamental characteristics of enterprise systems. A continuous performance capability is imposed in financial, communication, and other fields that use enterprise systems for applications such as stock exchange transaction handling, credit and debit card systems, telephone networks, and the like. Highly reliable systems are often implemented in applications with high financial or human costs, in circumstances of massive scaling, and in conditions that outages and data corruption cannot be tolerated.
Some systems combine multiple redundant processors running the same operations so that an error in a single processor can be detected and/or corrected. Results attained for each of the processors can be mutually compared. If all results are the same, all processors are presumed, with high probability of correctness, to be functioning properly. However, if results differ analysis is performed to determine which processor is operating incorrectly. Results from the multiple processors can be “voted” with the “winning” result determined to be correct. For example, a system with three processor elements typically uses the result attained by two of the three processors.
A difficulty arises for duplex systems with two executing processors since the even number of processor elements can result in a “voting tie” situation that may lead to aborted operation and outage. Ties can be avoided by running an odd number of processors, although a single processor does not have the fault detection capability provided by voting. A three or more processor system adds product cost.
In accordance with an embodiment of a redundant-processor computing device, an error handling method comprises detecting equivalent disparity among processor elements of the computing device operating and responding to the detected equivalent disparity by evaluating secondary considerations of processor fidelity.
Embodiments of the invention relating to both structure and method of operation may best be understood by referring to the following description and accompanying drawings:
A processor may incorporate multiple, redundant, loosely-coupled processor elements for error detection. A duplex arrangement using two processor elements is susceptible to a“voting tie” situation. Ties may be avoided by using an odd number of processors at the expense of fault detection capability if a single processor element is used and by adding costs for incorporation of additional processor elements. The illustrative system and method may use other information to resolve conflicts and break ties. Accordingly, an effective processor may be configured using only two processor elements for voting or comparison.
Referring to
The control element 106 determines whether evaluation of the secondary considerations is insufficient to resolve the disparity among the processor elements 104A, 104B and, if so, terminates computing device operations.
The computing device 102 can be a computer processor that uses multiple, redundant, loosely-synchronized processor elements 104A, 104B to detect and manage errors. A configuration with an even number of processor elements 104A, 104B is susceptible to a voting “tie” condition in which actions or results from the processor elements differ. For example, a computing device 102 may have two processor elements 104A, 104B so that any disparity is equivalent and results in a tie condition. Typically, an odd number of processing element, for example three, can be used at added cost to avoid ties.
In some situations, other information which may be called secondary considerations of fidelity may be available to resolve the disparity and break the tie. The other information is heuristic data which is sufficiently predictive to be trusted for disparity resolution. If the tie cannot be broken by use of the other information, then the error is considered sufficiently serious that the processor is halted due to an inability to guarantee correctness of either of the unequal voted data items.
Some embodiments may include a control element 106 that evaluates the secondary conditions of processor fidelity while the processor elements 104A, 104B are executing before equivalent disparity is detected and sets a probation vector 108 according to the evaluation. For example, the probation vector 108 may be implemented in a voting unit 110 and used by the voting unit 110 to resolve disparities and break ties in predetermined conditions. In a particular example, the probation vector 108 can have one bit of state per processor element 104A, 104B. Control logic, such as software, executing in each processor element 104A, 104B can set the bit in conditions that the logic has accumulated information for usage in breaking future ties, or very recent ties. The control logic can periodically reset the probation vector bits.
The voting unit 110, upon detecting a disparity or tie condition, may delay acting upon the condition or declaring a fatal error situation. Instead, the voting unit 110 can hold the compared values for a short duration time period before acting. Accordingly, the control element 106 can interject a delay between equivalent disparity detection and termination of computer device operation. The delay enables control logic, for example software, to possibly detect other errors or gather information pertinent to resolving the disparity or breaking the tie. The delay can also break potential race conditions. For example, if a self-detectable error occurs simultaneously or nearly the same time as a misvote, the delay enables further collection of information or analysis before the voter declares the misvote condition, enabling recognition of the error and resolution of the vote. In a particular duplex embodiment, if the control logic sets one of the two bits in the probation vector 108 during the short delay imposed by the voting unit 110, then the voting unit 110 resolves the disparity or breaks the tie in favor of the processor element, either 104A or 104B, that is not on probation. While the condition remains an error condition, the error is made recoverable. If the control logic does not set the bit in the probation vector prior to or during the delay, the error is considered to be fatal to the computing device 102, and operation is halted.
In a particular embodiment, the computing device 102 can include a logical synchronization unit 112 that contains the voting unit 110 and input/output interfaces 114A and 114B. For example, the interfaces 114A and 114B may include a programmed input/output (PIO) interface and a direct memory access (DMA) interface. A possible equivalent disparity or tie condition may include a condition of a first processor element that performs a programmed input/output (PIO) action while a second processor element does not. A second example of a equivalent disparity or tie condition may be a miscompare on voted data whereby data supplied by the two processor elements 104A, 104B does not match on a programmed input/output (PIO) action or a direct memory access (DMA) action. Other equivalent disparity or tie conditions include a first processor element performing a PIO read while a second processor element performs a PIO write, or first and second processors reading or writing different addresses.
Referring to
The probation vector 204 is used to monitor secondary considerations of processor element fidelity before an error is detected, when more abundant information relating to processor element conditions and functionality are available. In contrast, a system that does not begin acquiring status information until an error is detected may have more limited functional capabilities, and a possible inability to perform actions diagnostic of processor element fidelity. Acquisition of status and operational information while the processor elements 202A, 202B are executing in due course simplifies operation because information is merely noted when available. Higher complexity algorithms that execute following error or disparity detection and require information to be evoked or stimulated, as well as monitored, can be avoided.
The computing system 200 also includes a voter 208 that is coupled to the plurality of processor elements 202A, 202B and compares actions taken by the processor elements 202A, 202B to determine disparity in processor actions. A control element 210 responds to disparity among the processor elements 202A, 202B based on the probation vector 204. In some embodiment, the control element 210 can interject a delay between disparity detection and computer device operation termination to allow monitoring of additional information that may be useful in resolving disparity and breaking ties. The control element 210 can also determine whether evaluation of processor fidelity is insufficient to resolve the disparity among the processor elements 202A, 202B and, if so, terminate computing system operations.
Referring to
The selected delay has duration sufficient to enable near-simultaneous arrival of information for usage in resolving the disparity condition. The delay is imposed in case of processor disparity or tie, to enable simultaneous or near-simultaneous arrival of information that can be used in disparity resolution and/or tie-breaking. The delay has suitable duration to enable logic to receive a high priority interrupt and perform a few computations and is sufficiently long to enable information to be acquired at the same time or very close to occurrence of an error. Typical delay duration, using current processor operating speeds and technology, is on the order of tens or hundreds of microseconds, sufficient to handle the interrupt and execute hundreds or thousands of instructions. The delay may assist in avoiding or breaking race conditions.
The selected delay also has an upper limit. A lengthy examination of error state or diagnostic execution may not be acceptable. During the time between disparity and selection of the winning processor, the parallel input/output (PIO) operation or direct memory access (DMA) operation is suspended, possibly disrupting communications with other processors as well as communications on the interprocessor network due to backpressure. Such disruption is generally not desirable and is avoided by imposing an upper limit on the delay duration.
Referring to
A processor complex with one slice is called a simplex complex. A two-slice processor complex is called a duplex, dual modular redundant, or DMR complex. A three-slice processor complex is called a triplex, tri-modular redundant, or TMR complex. A processor complex 400 includes both processor elements 402 and corresponding logical synchronization units 412.
A computing system comprises one or more logical processors 408. The computing system also comprises one or more processor complexes 400. The processor complexes 400 are interconnected via a network, for example a System Area Network (SAN), a local area network (LAN), a wide area network (WAN), or the like, or simply a connection to a bus. The network is used for connection to both other processors and to input/output (I/O) devices. Voting or output data comparison is performed for all data transfers between the logical processor and the network or the network I/O adapter.
In a logical processor 408, one, two, three, or possibly more processor elements cooperate to perform logical processor operations. Cooperative actions include coordinating or synchronizing mutually among the processor elements, exchanging data, replicating input data, and voting on operations and output data selection. In the illustrative embodiment, the various cooperative actions can be implemented within or supported by implementation of the logical synchronization units 412.
Referring to
The computing system 500 may also include a logical synchronization unit 512 comprising the voter 504 and a network interface 520, such as a SAN interface that can be configured to connected to the network 518 by one or more ports. For example the network connection may be made as shown in
The logical synchronization unit 512 maintains multiple bit sets that represent, at any point in time, the set of processor elements that are enabled to perform selected operations. The voter 504 includes a plurality of multiple-bit configuration registers to indicate which processor elements are expected and enabled to participate in selected operations, for example including programmed input/output (PIO) operations with the processor elements 502A, 502B, and network interface 520 DMA operations including direct memory access (DMA) output voting and DMA input replication.
The voter configuration bits represent which processor elements are meant to be “assigned” to a logical processor and are therefore eligible for performing various operations such as output voting operations. Configured processor elements are defined, for any particular operation type, as the set of processor elements enabled by the configuration bits in the logical synchronization unit 512 to perform that operation type. The logical synchronization unit 512 ignores operations of non-configured processor elements on an operation-type basis. Configuration bits are set whenever processor elements “join” the configuration of a logical processor 508. Configuration bits are reset or cleared when processor elements leave the configuration, for example by being voted out.
Participating processor elements are the set of configured processor elements, for a particular instance of an operation, that actually perform the operation at a particular time, within a timeout period. The set of configured processor elements is related to a type of operation while the set of participating processor elements is related to an individual instance of an operation. The configured processor elements have outbound data transfers voted. Participating processor elements are the elements that actually issue a particular outbound data transfer.
For a particular voted operation, all configured processor elements are expected to participate. A processor element that does not participate times out on the operation.
Generally, voted operations can result in conditions including full agreement, timeout, simple miscompare, tie miscompare, and full miscompare. In full agreement, all configured processor elements participate within a reasonable time and supply identical data for an identical operation so that all voted data matches. In a timeout condition, one or more configured processor elements do not participate in time. For a simple miscompare, a strict majority of configured processor elements supplies identical data, and a strict minority supplies different data. A strict majority is greater than half, for example one in simplex, two in duplex, and two or three in triplex. A tie miscompare occurs with an even number of processor elements, for example two, configured in which the data does not compare. For a full miscompare, all sets of voted data, for example all three in triplex, miscompare pair-wise so that no strict majority results. The different types of conditions are mapped into one of three error conditions including no error, minor error, and major error.
For no error, the operation proceeds and no error is reported. For a minor error, the operation proceeds but an error is reported. Minor errors include such timeouts or disagreements among voted data in which available information is sufficient to enable resolution. Triplex configurations in which two processors are in agreement and duplex configurations in disagreement when other information is available to resolve the disagreement are examples of minor error conditions. For a major error, the operation does not proceed, but rather is aborted, and an error is reported. Major errors include such timeouts and disagreements among voted data that the condition cannot be resolved. Triplex configurations with three-way disagreement or timeouts in two of the three processors, and duplex configurations with disagreeing processors and no available information for resolution are examples of major error conditions.
In addition to timeouts and disagreement among redundant processors, other errors that may be handled using the illustrative techniques include breaks in cabling, for example between the processor elements and voter. Such errors can often be self-signaling. Self-signaling errors are defined as errors detected by an explicit detection element or mechanism as distinguished from implicit detection techniques such as voting. Various types of errors may be self-signaling. Examples of self-signaling errors include direct memory access (DMA) read timeouts, errors detected by parity or other error checking codes, loss-of-signal in optical signals, and loss of electrical continuity for electrical signals. With respect to the various illustrative systems, voting detects errors, but in a duplex configuration voting alone does not distinguish which voted data is correct and which incorrect. Self-signaling errors designate which of the two data suppliers is incorrect.
The processor elements 502A, 502B each have a memory 514 or share a partitioned memory with other processor elements in the same slice.
The PIO operations are processor-initiated reads (loads) or writes (stores) to any part of the processor's address space other than “main memory”. The address space may contain registers, pseudo-registers, and memory other than main memory. PIO operations may be targeted to resources in either the voter 504 or to a network interface 520. Operations targeted to the voter 504 may be either unvoted or voted (compared), depending on the address of the register being accessed. Operations to the network interface 520 may always be voted in some implementations.
When any configured processor element initiates a voted PIO read, the voter 504 captures the operation and read address and sets a timer. The voter 504 waits for all configured processor elements to perform the same operation. When all configured processor elements initiate the operation, the operation and address are compared, for example a bit-by-bit comparison of the entire operation and address. If one or more of the configured processor elements fails to initiate the operation within a configurable timeout period, a PIO timeout condition exists. In circumstances that the illustrative error handling and tie-breaking technique is not implemented or is disabled, operation can be described as follows. If all configured processor elements participate, then for the case of full agreement or simple miscompare, the operation proceeds, ignoring miscompared data, if any. A simple miscompare is handled as a minor error. Otherwise (not full agreement or simple miscompare), the operation is aborted—not performed, and a “bus error” is returned to all requesting processor elements, and a major error is reported. If the operation is not aborted, then if the PIO read is targeted to the network interface 520, the voter 504 forwards the operation and address to the network interface 520 and waits for a response. If the operation is targeted to the voter 504, then the voter 504 accesses data directly. The response data, when available, is replicated by the voter 504 and sent as a response to all participating processor elements at approximately the same instant.
When any configured processor element initiates a voted PIO write operation, the voter 504 captures the operation, write address, and write data, then sets a timer and waits for all configured processor elements to perform the same operation. When all configured processor elements initiate the operation, the operation, write address, and write data are bit-by-bit compared. If one or more of the configured processor elements fails to initiate the operation within a configurable timeout period, a PIO timeout condition exists. In circumstances that the illustrative error handling and tie-breaking technique is not implemented or is disabled, the operation is as follows. If all configured processor elements participate, in the case of full agreement or simple miscompare, the operation proceeds, ignoring miscompared data. A simple miscompare is handled as a minor error. Otherwise (not full agreement or simple miscompare), the operation is aborted—not performed. No direct response is made to the processor element because no response or acknowledgement is normally made to write operations. When the operation is aborted, a major error is reported. An operation that is not aborted is handled according to the target address of the write operation. For a PIO write to the network interface 520, the voter 504 forwards the operation, address, and data to the network interface 520. For a PIO write to the voter 504, the voter 504 performs the write operation directly. No response is made to the processor element. Because of the possible side-effects with PIO write operations, and because software does not necessarily verify the effect, or success, of each write operation, when a PIO write operation is aborted due to a major voting error, the voter 504 suspends all future PIO write operations, which are also aborted, until the software detects the error and re-enables PIO voting. Typically the error is detected by handling an error interrupt. Note, however, that in the triplex case for a simple miscompare such as one processor element write and two processor elements time out, no requirement is made to abort all future programmed I/O write operations.
PIO operations are initiated by the processor elements, as contrasted to DMA operations which are initiated by the network interface 520. Therefore, PIO timeouts are possible in two different circumstances. In a first circumstance, one or two processor elements, operating correctly, initiate a PIO operation, and one processor element, operating incorrectly or stopped, fails to perform the PIO operation. The error is detected when the timer expires. Without further information, the processor element operating incorrectly may be indeterminable, for example when two processor elements are configured and one times out.
In a second circumstance, one processor element, operating incorrectly, initiates a PIO operation that should not occur, and the other processor element or processor elements, operating correctly, do not initiate a PIO operation. Again the error is detected when the timer expires, although without further information the processor element operating incorrectly may be indeterminable, for example for two active processor elements, one of which times out. Accordingly, the processor elements that do not participate are not necessarily incorrect.
In the triplex case, the strict majority is always trustworthy so that if one processor element times out and the remaining two processor elements perform a PIO operation, then the processor element that times out is in error and ignored, while the PIO operation proceeds and the data voted with a minor error indicated. Also in the triplex case, if two processor elements time out when a single processor element performs a PIO operation, then the processor element that performs the PIO operation is in error and the PIO is ignored. The lone processor element can be considered a “rogue processor element” and the attempted PIO operation is called a “rogue operation”.
In the duplex case, whether the processor element performing the PIO or the processor element that is not participating is in error cannot be determined, without other evidence. Accordingly a tie or disparity condition exists.
Some embodiments of the computing system 500 further comprise a two-processor elements configuration 502A, 502B, the programmed input/output (PIO) interface 522, and a direct memory access (DMA) interface 524 coupled to the voter 504. An action disparity that is detectable by the control element 506 is a miscompare on voted data with non-matching data supplied by two processor elements 502A, 502B either on a PIO action or a DMA action.
Direct memory access (DMA) reads are outbound operations initiated by the network interface 520. The voter 504 replicates and forwards the DMA read operation and address to all configured processor elements at approximately the same instant, subject to congestion delays in the different slices that may cause an operation to arrive at the processor elements at slightly different times. The voter 504 then starts a timer and waits for the responses. Response data flows from the processor elements 502A, 502B, through the voter 504, to the network interface 520. Responses arriving from the configured processor elements are buffered in the voter 504 rather than being sent immediately to the network interface 520. The later responses, upon arrival, are compared with the earlier responses saved in the data buffers. When a strict majority of the responses agree, one copy of the data, from one of the agreeing responses, is communicated to the network interface 520. If a strict majority of the responses do not agree, then data is not sent to the network interface 520 in a manner that can be interpreted as valid data.
Any processor element that does not respond within the timeout period is declared to have timed out. Unlike PIO timeouts, no rogue situations can occur with DMA timeouts because the DMA operation is initiated through the network interface 520. Accordingly, if one processor element times out, that processor element is necessarily erroneous in both duplex and triplex cases. The condition is considered a minor error, and the DMA operation proceeds using data supplied by the processor element or processor elements that do not time out. If two processor elements time out, or the only processor element in a simplex case, then a major error is indicated and the DMA operation is aborted. The voter 504 generates an error notification interrupt to all configured processor elements in the case of any disagreement or timeout, whether data is successfully forwarded to the network interface 520 or otherwise. The interrupt indicates which processor elements did time out, if any, and all comparison results.
Direct memory access (DMA) writes are inbound operations with data flowing from the network interface 520 through the voter 504 to the processor elements 502A, 502B. DMA write operations are initiated by the network interface 520. The voter 504 replicates and forwards the operation, address, and data to all configured slices at approximately the same instant. No response is made to the network interface 520.
The voted PIO operations and DMA responses are protected by timeouts. PIO and DMA timeout values are both configurable by control logic, such as software, and may be different.
For PIO timeouts, the timer is started when the first PIO operation arrives. The timer is restarted when each operation arrives, giving a full timeout period for the later arrivals relative to the earlier ones, a behavior used in all implementations due to the possibility of a PIO operation being initiated by a “rogue processor element”. PIO operations can never time out in the simplex case because the operation originates from the processor element.
For DMA read response timeouts, DMA operations can time out in simplex, duplex, and triplex configurations because the operation is originated from the network interface 520. The timer is started when the DMA request is forwarded by the voter 504 to memories of all processor elements. The timer may optionally be restarted when each response arrives, giving a full timeout period for the later responses. Alternatively, a single timeout interval may be applied to all configured processor elements. Either option is possible since no rogue DMA operations can occur.
In an illustrative embodiment, special case handling can be used when a PIO timeout occurs or a miscompare occurs on voted data in a duplex configuration. In the illustrative example, duplex tie handling generally is inapplicable to DMA timeouts, and to simplex or triplex configurations. Two disparity or tie conditions include a PIO timeout in which one processor element performs a PIO operation and the other processor element does not, and a miscompare on voted data in which data supplied by the two processor elements does not match, either on a PIO operation or a DMA read operation.
In the absence of any other information, a tie or disparity condition in a duplex configuration is ambiguous whereby the trustworthiness of each processor element is not obvious, leading to a typical policy of halting the logical processor.
However, occasionally, other information, termed secondary considerations of processor fidelity, may exist. The other information may be considered sufficiently strong, even if circumstantial, to implicate one of the two processor elements. For example, a recent logged history of other detected recoverable errors may be indicative of a degradation of processor element reliability.
Another example of pertinent reliability information is a recent history of processor replacement. A newly-reintegrated processor element or a new slice may be a more likely source of error than an element that has long been installed without a history of error. Such early life problems are frequently discovered within a short time, on the order of minutes, following installation.
A further example of reliability information is inherent in the multiple-dimensional configuration of logical and physical processors, for example as shown in
Some embodiments of the computing system 500 further comprise a probation vector 526 coupled to the voter 504 and coupled to the processor elements 502A, 502B. The probation vector 526 holds a signal for each processor element 502A, 502B. A processor control policy executable on the processor elements 502A, 502B evaluates the secondary conditions of processor fidelity during processor element execution. The processor control policy sets bits in the probation vector according to the evaluation of secondary considerations before determining whether a major or minor error has occurred. In a particular embodiment, the probation vector 526 enables implementations to supply the other information to the voter 504 in such a way that duplex tie conditions can be simply resolved. The probation vector 526 comprises one bit for each processor element 502A, 502B. The logical synchronization unit 512 uses the probation vector 526 as a tie-breaker in some conditions.
In an illustrative embodiment, PIO writes to the probation vector 526 are not voted. The initial or reset value of the probation vector 526 is all zero. Each processor element 502A, 502B can set or reset the probation bit only for the processor slice associated with the processor element, and not for any other slice. The probation vector 526 is not used in simplex and triplex modes for which the bits may still be set by the control logic, but are ignored.
The probation vector 526 is ignored if both processor slices agree, indicating no error. The probation vector 526 is also ignored in the case of self-signaling errors, which indicate the error source so that tie-breaking is superfluous. Self-signaling errors are thus classified as minor errors. The errors occur when the voter 504 can be certain that one slice should be ignored, for example in an outbound DMA read response, when one slice does not respond and times out, or when the slice supplies data marked as “known bad”. “Known bad” data relates to another example of self-signaling error and is data returned from a processor element's memory or cache that generates a detected error, such as a parity or other detected error, when accessed.
Control logic in each processor element sets the associated probation bit independently of the other processor elements; reaching an agreement among processor elements is not required. Accordingly, both probation bits may be asserted at any time. The control logic resets the probation bits after some amount of time. The time duration is an implementation-defined parameter or policy.
In some circumstances, the probation bits may be set for all processor elements on a processor slice due to an error that places behavior of the entire slice in doubt. Accordingly, the control logic can propagate the probation bits from one processor element, where an error has been detected, to other processor elements in the slice by an implementation-defined technique. Examples of an implementation-defined system include exchanging probation bits via a register in a slice application-specific integrated circuit (ASIC), and/or using inter-processor, intra-slice interrupts.
In some embodiments of the computing system 500, the control element 506 interjects a delay between equivalent disparity detection and computing device operation termination. When a duplex, non-self-signaling error occurs, a delay is imposed prior to declaring the situation an error condition. During the delay period, operation is held in limbo. The operation is neither completed nor aborted. The delay enables the tie to be broken by control logic setting the probation bit after the error but before the timeout elapses. The delay does not occur, and therefore does not add any latency, in full agreement cases, and in simplex and triplex cases.
The delay period in a tie-breaker situation is maintained sufficiently small to avoid excessive network backpressure, therefore preventing an increase in congestion in the network. For current technology, a sufficiently small delay may be of the order of the range of tens or hundreds of microseconds.
In an illustrative embodiment, when the delay period begins, the logical synchronization unit 512 sends an interrupt to all participating slices to inform the slices that a miscomparison has occurred, although an error has not been declared. The interrupt is in addition to the interrupt that is generated at the end of the delay, when the voting error is final, either major or minor.
In a tie-breaker or disparity condition, if after the delay interval has elapsed, both probation bits are enabled or both probation bits are disabled, then a major error exists and the operation is aborted. If the bits are in opposite states, then the slice with the probation bit off or reset is obeyed, and the slice with the probation bit is on or set is ignored. If the probation bits are used to break the tie, then a minor error is declared.
The policy for setting probation bits and duration that the bit setting is maintained is implementation-specific.
In the illustrative embodiment, the logical synchronization unit 512 reports all errors, both major and minor, to control logic, such as software via an interrupt and status register. In one implementation, status register bits indicate that the tie-break mechanism has been invoked and designate which processor element or slice is obeyed.
Referring to
The method can further comprise determining whether evaluation 624, 634 of the secondary considerations is insufficient to resolve the disparity among the processor elements. If so, computing device operations are terminated 626 and 628, 636 and 638. Delay can be inserted 620 and 622, 630 and 632 between equivalent disparity detection and computer device operation termination 626, 636.
In the illustrative method 600, control logic receives 602 a programmed input/output (PIO) request from one processor element (PE), buffers 604 the request, and starts 606 a timer. If a second request is received 608, the two requests are compared 610. Otherwise, if the timer has not elapsed 612, whether the second request is received is determined 608. If the timer has elapsed 612, analysis of secondary considerations of processor fidelity begins.
In conditions that a second request is received 608, following comparison 610 of the requests, if the requests match 614, the control logic performs the PIO operation 616 and no error is indicated, and the method is complete 618. Otherwise, an equivalent disparity condition exists in the form of a miscompare on voted data whereby command, address, or data supplied by two processor elements does not match on a programmed input/output (PIO) action. Operation, address, and, for a write operation, data are compared to determine a match condition. Secondary consideration analysis begins with the control logic sending 620 a “tie break pending” interrupt to the processor elements, and waiting 622 the configured time. Generally, the secondary conditions of processor fidelity can be evaluated during processor element execution, and a probation vector can be set according to the evaluation prior to determination of major or minor error. If probation bits are equal 624, the operation is aborted 626 since the tie or disparity condition cannot be resolved and a major error interrupt is sent to the processor elements. The method is terminated 628.
Otherwise, if the probation bits are not equal 624, the control logic follows direction 648 of the processor element that is not on probation, and the PIO operation specified by the non-probation processor element is performed. A minor error interrupt is sent 650 to the processor elements, and the method completes 652 with a suitable minor error handling technique. In some embodiments, the minor error is addressed by marking the loser of the voting decision as no longer participating in the logical processor. Subsequent input/output operations or other accesses to the voter are ignored. In other embodiments, software processing in the loser is interrupted and software executing in remaining processor elements shuts down the offending processor element.
For analysis of secondary considerations of processor fidelity after the timer elapses 612, an equivalent disparity condition occurs in which a first processor element performs a programmed input/output (PIO) action while a second processor element does not. Control logic sends 630 a “tie break pending” interrupt to the processor elements, and waits 632 the configured time. If probation bits are equal 634, the operation is aborted 636 since correct operation cannot be determined. A major error interrupt is sent to the processor elements. The method is terminated 638.
Otherwise, if the probation bits are not equal 634, control logic determines whether the processor element requesting the PIO operation is on probation 640. If the processor element is on probation 640, the control logic follows direction 642 of the processor element that is not on probation and ignores or aborts the PIO operation, sends 644 a minor error interrupt to the processor elements, handles the minor error, and terminates 646 the method. If the processor element requesting the PIO is not on probation 640, the control logic follows direction 648 of the processor element that is not on probation, and the PIO operation specified by the non-probation processor element is performed. A minor error interrupt is sent 650 to the processor elements, and the method is complete 652.
Referring to
The method can further comprise determining whether evaluation 738 of the secondary considerations is insufficient to resolve the disparity among the processor elements. If so, computing device operations are terminated 740 and 742. Delay can be inserted 736 between equivalent disparity detection and computer device operation termination 740, 742.
In the illustrative method 700, control logic receives 702 a direct memory access (DMA) read request from a network agent such as a system area network (SAN) agent, replicates and forwards 704 the request to both processor elements, and starts 706 a timer. If a first response is received 708, the first response is buffered 716. In some embodiments, the timer may be restarted as the first response is buffered 716. If the timer has not timed out 710, whether the first response is received is again determined 708. If the timer has timed out 710, the operation is aborted 712 and a major error interrupt is sent to both processor elements. The major error interrupt is indicative of a double timeout. The method is then terminated 714.
In conditions that the first response is received 708 and the first response is buffered 716, whether a second response has been received is determined 718. If the second response has been received 718, the first and second response data are compared 726. Otherwise, if the timer has not timed out 720, whether the second response has been received is again determined 718. If the timer has timed out 720, the operation is completed 722 using data from the first response and a minor error interrupt is sent to both processor elements. The single timeout condition is indicative of a “self-signaling error”. The method is then terminated 724.
In conditions that the second response is received 718 and data from the first and second responses is compared 726, if the first and second response data are equal 728 the data match so that the operation is completed 730 with no error. The method completes successfully 732. Otherwise, the first and second response data are unequal 728 and a “tie break pending” interrupt 734 is sent to the processor elements. A delay is inserted 736 to wait for a configured time. Probation bits are read to determine whether the probation bits are equal 738. If so, the operation is aborted 740 since the tie cannot be resolved using the probation bits and a major error interrupt is sent to the processor elements. The method terminates unsuccessfully 742. Otherwise, the probation bits are not equal 738 and the operation is completed 744 using the response from the processor element that is not on probation. A minor error interrupt is sent 746 to the processor elements, the minor error handled by marking the loser for removal or shutting down the offending processor element, and the method is terminated 748.
While the present disclosure describes various embodiments, these embodiments are to be understood as illustrative and do not limit the claim scope. Many variations, modifications, additions and improvements of the described embodiments are possible. For example, those having ordinary skill in the art will readily implement the steps necessary to provide the structures and methods disclosed herein, and will understand that the process parameters, materials, and dimensions are given by way of example only. The parameters, materials, and dimensions can be varied to achieve the desired structure as well as modifications, which are within the scope of the claims. For example, although the illustrative structures and methods are most applicable to multiple-processor systems in a duplex configuration, various aspects may be implemented in configurations with more or fewer processors. Furthermore, the illustrative embodiments depict particular arrangements of components. Any suitable arrangement of components may be used. The various operations performed may be implemented in any suitable matter, for example in hardware, software, firmware, or the like.
| Number | Date | Country | |
|---|---|---|---|
| 60557812 | Mar 2004 | US |
