Sangeetha Seshadri*, Lawrence Chiu†, Cornel Constantinescu†, Subashini Balachandran†, Clem Dickey†, Ling Liu* and Paul Muench†
|*Georgia Institute of Technology|
801 Atlantic Drive
Atlanta, GA - 30332
|†IBM Almaden Research Center|
650 Harry Road
San Jose, CA - 95120
With software failures and bugs becoming an accepted fact, focusing on recovery and reducing time to recovery has become essential in many modern storage systems today. In current system architectures, even with redundant controllers, most microcode failures trigger system-wide recovery [9,10] causing the system to lose availability for at least a few seconds, and then wait for higher layers to redrive the operation. This unavailability is visible to customers as service outage and will only increase as the platform continues to grow using the legacy architecture.
In order to reduce the recovery time and more importantly scale the recovery process with system growth, it is essential to perform recovery at a fine-grained level: recovering only failed components and allowing the rest of the system to function uninterrupted. However, due to fuzzy component interfaces, complex dependencies and involved operational semantics of the system, implementing such fine-grained recovery is challenging. Therefore, firstly we must develop a mechanism to perform fine-grained recovery taking into consideration interactions between components and recovery semantics. Secondly, since localized recovery spans multiple dependent threads in reality, we must bound this localized recovery process in time and resource consumption in order to ensure that resources are available for other normally operating tasks even during recovery. Finally, since we are dealing with a large legacy architecture (> 2M loc), to ensure feasibility in terms of development time and cost we should minimize changes to the architecture.
In this paper we develop a recovery conscious framework for multi-core architectures and a suite of techniques for improving the resiliency and recovery efficiency of highly concurrent embedded storage software systems. Our techniques aim at providing continuous availability and good performance even during a recovery process by bounding the time to recovery while minimizing the need for rearchitecting the system.
The main contributions of our recovery conscious framework include (1) a task-level recovery model, which consists of mechanisms for classifying storage tasks into recovery groups and dividing the overall system resources into recovery-oriented resource pools; and (2) the development of recovery-conscious scheduling, which enforces some serializability of failure-dependent tasks in order to reduce the ripple effect of software failures and improve the availability of the system. We present three alternative recovery-conscious scheduling algorithms, each representing one way to trade-off between recovery time and system performance.
We have implemented and evaluated these recovery-conscious scheduling algorithms on a real industry-standard storage system. Our experimental evaluation results show that the proposed recovery conscious scheduling algorithms are non-intrusive, involve minimal new code and can significantly improve performance during failure recovery thereby enhancing availability.
Figure 1 gives a conceptual representation of a storage subsystem. This is a single storage subsystem node consisting of hosts, devices, a processor complex and the interconnects. In practice, storage systems may be composed of one or more such nodes in order to avoid single-points-of-failure. The processor complex provides the management functionalities for the storage subsystem. The system memory available within the processor complex serves as program memory and may also serve as the data cache. The memory is accessible to all the processors within the complex and holds the job queues through which functional components dispatch work. As shown in Figure 1, this processor complex has a single job queue and is an N-way SMP node. Any of the N processors may execute the jobs available in the queue. Some storage systems may have more than one job queue (e.g. multiple priority queues).
The storage controller software typically consists of a number of interacting components each of which performs work through a large number of asynchronous, short-running threads (~ μ secs). We refer to each of these threads as a `task'. Examples of components include SCSI command processor, cache manager and device manager. Tasks (e.g., processing a SCSI command, reading data into cache memory, discarding data from cache etc.) are enqueued onto the job queues by the components and then dispatched to run on one of the many available processors each of which runs an independent scheduler. Tasks interact both through shared data-structures in memory as well as through message passing.
With this architecture, when one thread encounters an exception that causes the system to enter an unknown or incorrect state, the common way to return the system to an acceptable, functional state is by restarting and reinitializing the entire system. Since the system state may either be lost, or cannot be trusted to be consistent, some higher layer must now redrive operations after the system has performed basic consistency checks of non-volatile metadata and data. While the system reinitializes and waits for the operations to be redriven by a host, access to the system is lost contributing to the downtime. This recovery process is widely recognized as a barrier to achieving high(er) availability. Moreover, as the system scales to larger number of cores and as the size of the in-memory structures increase, such system-wide recovery will no longer scale.
The necessity to embark on system-wide recovery to deal with software failures is mainly due to the complex interactions between the tasks which may belong to different components. Due to the high volume of tasks (more than 20 million/minute in a typical workload), their short-running nature and the involved semantics of each task, it becomes infeasible to maintain logs or perform database-style recovery actions in the presence of software failures. Often such software failures need to be explicitly handled by the developer. However, the number of scenarios are so large, especially in embedded systems, that the programmer cannot realistically anticipate every possible failure. Also, an individual developer may only be aware of the clean-up routines for the limited scope being handled by them. This knowledge is insufficient to recover the entire system from failures, given that often interactions among tasks and execution paths are determined dynamically.
Many software systems, especially legacy systems, do not satisfy the conditions outlined as essential for micro-rebootable software . For instance, even though the storage software may be reasonably modular, component boundaries, if they exist, are very loosely defined. In addition, the scenario where components are stateful and interact with other components through globally shared structures (data-structures, metadata), often leads to components modifying each other's state irreversibly. Moreover, resources such as hardware and software locks, devices and metadata are shared across components. Under these circumstances, the scope of a recovery action is not limited to a single component.
The discussion above highlights some key problems that need to be addressed in order to improve system availability and provide scalable recovery from software failures. Concretely, we must answer the following questions:
Our work is targeted at dealing with such transient failures in a storage software system and in particular the embedded storage controller's microcode. Below, we provide a classification of transient failures which we intend to deal with through localized recovery.
In complex systems, often code paths are dynamic and input parameters are determined at runtime. As a result many faults are not caught at compile time. On pure functions, faults may be classified as:
With actions based on system state there are additional complexities. For example, a configuration issue that appeared early in the installation process may have been fixed by trying various combinations of actions that were not correctly undone. As a result the system finds itself in an unknown state which manifests as a failure after some period of normal operation. Such errors are very difficult to trace, and although transient may continue to appear every so often. We classify such system state based errors as:
Each of the above error types can lead to transient failures. Some of the transient failures can be fixed through appropriate recovery actions which may range from dropping the current request to retrying the operation or performing a set of actions that take the system to a known consistent state. For example, some of such transient faults that occur in storage controller code are:
In production environments, techniques for fault-tolerance, i.e., coping with the existence and manifestation of software faults can be classified into two primary categories with respect to the fault repairing methods: (1) those that provide fault treatment, such as restarts of the software, rebooting of the system and utilizing process pair redundancy; and (2) those that provide error recovery, such as check-pointing and log-based recovery. Alternatively, one can categorize the recovery models based on the granularity of the recovery scopes. All the above-mentioned techniques could be applied to any recovery scope. In our context, we consider the following three types of recovery scopes:
The main advantage of performing error recovery or fault-treatment at the task-level as compared to the component-level, is that it allows us to accommodate cross-component interactions and define `recovery boundaries' in place of `component boundaries'. Our goal is to handle most of the failures and exceptions through task-level (localized) recovery, and avoid resorting to system-wide recovery unless it is absolutely necessary.
Recovery from a software failure involves choosing an appropriate strategy to treat/recover from the failure. The choice of recovery strategy depends on the nature of the task, the context of the failure, and the type of failure. For example, within a single system, the recovery strategy could range from continuing the operations (ignoring the error), retrying the operation (fault treatment using environmental diversity) or propagating the fault to a higher layer. In general, with every failure context and type, we could associate a recovery action. In addition, to ensure that the system will return to a consistent state, we must also avoid deadlock or resource hold-up situations by relinquishing resources such as hardware or software locks, devices or data sets that are in the possession of the task.
Bearing these design principles in mind, we develop a two-tier framework for performing task-level recovery through a set of non-intrusive recovery strategies. The top tier provides the capabilities of defining the recovery scope at task level through a careful combination of both the programmer's specification at much coarser granularity and the system-determined recovery scope at finer-granularity. The bottom tier provides the recovery-conscious scheduling algorithms that balance the performance and the recovery efficiency.
In this framework, we refer to the context of a failure as a recovery point and provide mechanisms for developers to define clean-up blocks which are recovery procedures and re-drive strategies. A clean-up block is associated with a recovery point and encapsulates failure codes, the associated recovery actions, and resource information. The specification of the actual recovery actions in each of the clean-up blocks is left to the developers due to their task-specific semantics.
In our implementation, the recovery-conscious scheduler alone was implemented in approximately 1000 lines of code. A naive coding and the design effort for task level recovery would be directly proportional to the number of ``panics'' or failures in the code that are intended to be handled using our framework. In general, the coding effort for a single recovery action is small and is estimated to be around a few tens of lines of code (using semicolons as the definition of lines of code) per recovery action on average. Note that, the clean-up block does not involve any logging or complex book-keeping and is intended to be light-weight. A more efficient handling of clean-up blocks would involve classifying common error/failure situations and then addressing the handling of the errors in a hierarchical fashion. For example, recoveries may be nested and we could re-throw an error and recover with the next higher clean-up block defined in the stack. This would involve design effort towards the classification of error codes into classes and sub-classes and identification of common error handling situations. Finally, if we are unable to address an error using our framework, existing error handling mechanisms would be used as default. The point of recovery in the stack may be determined by factors such as access to data structures and possibilities of recovery strategies such as retrying, termination or ignoring the error.
Figure 2 shows a schematic diagram of the recovery framework using the call stack of a single task. As the task moves through its execution path, it passes through multiple recovery points and accumulates clean-up blocks. When the task leaves a context, the clean-up actions associated with the context go out of scope. On the other hand, nesting of contexts results in the nesting of the corresponding clean-up blocks and the framework keeps track of necessary clean-up blocks.
The clean-up blocks are gathered and carried along during task execution but are not invoked unless a failure occurs. Resource information can also be gathered passively. Such a framework allows a choice of recovery strategy based on task requirements and requires minimal rearchitecting of the system.
Example : We describe the selection of recovery strategy and design of clean-up blocks using an example from our storage controller implementation. Consider the error described in Figure 2 which depicts relevant portions of the call stack. The failure situation described in this example is similar to the commonly used `assert' programming construct. The error is encountered when a task has run out of a temporary cache data structure known as a `control block' which is not expected to occur normally and hence results in a `panic'.
In this particular situation, ignoring the error is not a possible recovery strategy since the task would be unable to complete until a control block is available. One possible strategy is to search the list of control blocks to identify any instances that are not currently in use, but have not been freed correctly (for example, due to incorrect flags). If any such instances exist, they could be made available to the stalled task. An alternative strategy would be to retry the operation beginning at the `WriteToCache' routine at a later time in order to work around concurrency issues. Retrying the operation may involve rolling back the resource and state setup along this call path to their original state. Resource blocks are used to carry the information required to successfully execute this strategy. Finally, in the case of less critical tasks, aborting the task may also be an option. Alternatively, consider a situation where an error is encountered due to a component releasing access to a track to which it did not have access in the first place. The error was caused due to a mismatch in the number of active users as perceived by the component. In this case, a possible recovery strategy would be to correctly set the count for the number of active users and proceed with the execution, effectively ignoring the error.
Note that, it is important we ensure that the interfaces with the recovery code and the recovery code itself are reliable. The most important issue in the tier one design is how to adequately identify the recovery scopes or boundaries, and how to concretely determine what are the set of tasks that need to undergo recovery upon a failure?
Dependencies between tasks may be explicit as in the case of functional dependencies or implicit such as those arising from shared access to stateful structures (e.g., data structures, metadata) or devices. For example, tasks belonging to the same user request may be identified as having the same recovery scope. Likewise, identical tasks belonging to the same functional component may also be marked with the same recovery scope. Explicit dependencies can be specified by the programmer.
However, explicit dependencies specified by the programmer may be very coarse. For example, an `adapter queue full' error should only affect tasks attempting to write to that adapter and should not initiate recovery across the component. Likewise, some dependencies may have been overlooked due to their dynamic nature and the immense complexity of the system. Therefore one way to refine explicit dependencies is to identify implicit dependencies continuously and utilize them to refine the developer-defined recovery scopes over time. For example, one approach to identifying implicit dependencies at runtime is by observing patterns of hardware and software lock acquisitions. We can group the tasks that share locks into the same recovery scope, since sharing locks typically implies that they have shared access to a stateful global object. Figure 3 illustrates this approach through an example. It shows five tasks and their respective lock acquisition patterns. Tasks T1, T2 and T4 are accessing overlapping sets of locks during their execution and thus are grouped into one recovery scope. Similarly, tasks T3 and T5 are grouped into another recovery scope. Clearly, this approach further refines the developer-specified recovery scope at task level into smaller recovery scopes based on runtime characterization of the dependencies with respect to lock acquisition.
Due to the space limit of this paper, we will omit the detailed development of recovery scope refinement mechanisms. In the remaining part of this paper, we assume that tasks are organized into disjoint recovery scopes refined based on implicit dependencies identified dynamically at runtime. In addition to recovery scopes, we argue that recovery-conscious resource management can be critical for improving system availability during localized recovery.
An important goal for providing fine-grained recovery (task or component level) is to improve recoverability and make efficient use of resources on the multi-core architectures. This ensures that resources are available for normal system operation in spite of some localized recovery being underway and that the recovery process is bounded both in time and in resource consumption. Without careful design, it is possible that more dependent tasks are dispatched before a recovery process can complete, resulting in an expansion of the recovery scope or an inconsistent system state. This problem is aggravated by the fact that recovery takes orders of magnitude longer (ranging from milliseconds to seconds) compared to normal operation (~ μ secs). Also a dangerous situation may arise where it is possible that many or all of the threads that are concurrently executing are dependent, especially since tasks often arrive in batches. Then the recovery process could consume all system resources essentially stalling the entire system.
Ideally we would like to ``fence'' the failed and recovering tasks until the recovery is complete. In order to do so we must control the number of dependent tasks that are scheduled concurrently, both during normal operation and during recovery. In the next section we discuss how to design a recovery conscious scheduler that can control how many dependent tasks are dispatched concurrently and what measures should be taken in the event of a failure.
In order to deal with the problem illustrated in Figure 4, we infuse ``recovery consciousness'' into the scheduler. Our recovery-conscious scheduler will enforce some serialization of dependent tasks thereby controlling the extent of a localized recovery operation that may occur at any time. To formally describe recovery conscious scheduling, we first define two important concepts: recovery groups and resource pools.
Recovery Groups: A recovery group is defined as the unit of a localized recovery operation i.e., the set of tasks that will undergo recovery concurrently. When clean-up procedures are initiated for any task within a recovery group, all other tasks belonging to the same recovery group that are executing concurrently will also initiate appropriate clean-up procedures in order to maintain the system in a consistent state. Note that recovery groups are determined based on explicitly specified dependencies that are further refined by the system based on observations of implicit dependencies. By definition, every task belongs to a single recovery group. Thus tasks in the system can be partitioned into multiple disjoint recovery groups.
Resource Pools: The concept of resource pools is used as a method to partition the overall processing resources into smaller independent resource units, called resource pools. Although we restrict resource pools in this paper to processors, the concept can be extended to any pool of identical resources such as replicas of metadata or data. Recovery conscious scheduling maps resource pools to recovery groups, thereby confining a recovery operation to the resources available within the resource pool assigned to it.
Proactive RCS comes into play during normal operation and enhances availability by enforcing some degree of serialization of dependent tasks. The goal of proactive scheduling is to reduce the impact of a failure by trying to bound the number of outstanding tasks per recovery group. Then in the event of a failure within any recovery group, the number of tasks belonging to that recovery group that are currently executing and need to undergo recovery are also controlled. By limiting the extent of a recovery process, proactive scheduling can help the system recover sooner, and at the same time, it controls the amount of resources dedicated to the recovery process. Proactive RCS thereby ensures resource availability to normal operation even during a localized recovery process.
The reactive aspect of recovery conscious scheduling takes over after a failure has occurred. When localized recovery is in progress, reactive RCS suspends the dispatch of tasks belonging to the group undergoing recovery until the recovery completes. This ensures quick completion of recovery by preventing transitive expansion of the recovery scope and avoiding deadlocks.
Note that our scheduling algorithms are concerned with partitioning resources between tasks belonging to different ``components'' of the same system which adds a second orthogonal level to the scheduling problem. We continue to respect the QoS or priority considerations specified by the designer at the level of user requests. For example, Figure 6 shows an existing QoS based scheduler using high, medium and low priority queues. Figure 7 shows how recovery-conscious scheduling used by a pool ρ1 dispatches jobs based on both priority and recovery-consciousness (by picking jobs only from the recovery groups assigned to it).
We use good-path and bad-path performance as the two main metrics for comparison of the recovery-conscious schedulers with performance oriented schedulers. By `good-path' performance we mean the performance of the system during normal operation. We use the term `bad-path' performance to refer to the performance of the system under localized failure/recovery.
Both good path and bad path performance can be measured using end-to-end performance metrics such as throughput and response time. In addition, we can also measure the performance of a scheduler from system-level factors, including cpu utilization, number of tasks dispatched over time, queue lengths, the overall utilization of other resources such as memory, and the ability to meet service level agreements and QoS specifications.
In order to provide a better understanding of the design philosophy of our recovery-conscious scheduling, we devise a running example scenario that is used to illustrate the design of all three classes of RCS algorithms. This running example has five resource pools: ρ1, ρ2, ρ3, ρ4 and ρ5 and four recovery groups: γ1, γ2, γ3 and γ4. We use σi to denote the recoverability constraint for the recovery group γi. Constraint σi specifies the upper limit on the amount of resources (processors in this case) that can be dedicated to the recovery group γi (1 ≤ i ≤ 4 in our running example). Since we are concerned with processing resources in this paper, it also indicates the number of tasks belonging to a recovery group that can be dispatched concurrently. The recoverability constraint σi is determined based on both the recovery group workload i.e., the number of tasks dispatched, and the observed task-level recovery time. Although recoverability constraints are specified from the availability standpoint, they must take performance requirements into consideration in order to be acceptable. Recoverability constraints are primarily used for proactive RCS.
For ease of exposition we assume that all resource pools are of equal size (1 processor each).
Table 1 shows the workload distribution between the recovery groups
and the recoverability constraint per group, where two processors are assigned to the recovery
group γ1 and one processor is assigned to each of the remaining three groups.
In contrast to the scenario in Table 1 where no resource pools are shared by two or more recovery groups, when more than one recovery group is mapped to a resource pool the scheduler must ensure that the dispatching scheme does not result in starvation. By avoiding starvation, it ensures that the functional interactions between the components are not disrupted. For example in our implementation we used a simple round-robin scheme for each scheduler to choose the next task from different recovery groups sharing the same resource pool. Other schemes such as those based on queue lengths or task arrival time are also appropriate as long as they avoid starvation.
This approach aims at achieving strict recovery isolation. As a result, it loses out on utilization of resources, which in turn impacts both throughput and response time. Although this is a naive approach to performing recovery-conscious scheduling it helps us in understanding issues related to the performance and recoverability trade-off. Note that all our RCS algorithms avoid starvation by using a round-robin scheme to cycle between recovery groups sharing the same resource pool. In systems where the workload is well understood and sparse in terms of resource utilization, static mappings offer a simple means of achieving serialization of recovery dependent tasks.
Implementation Considerations: There are two main data structures that are common to all RCS algorithms: (1) the mapping tables and (2) the job queues. Mapping table implementations keep track of the list of recovery groups assigned to each resource pool. They also keep track of groups that are currently undergoing recovery for the purpose of reactive scheduling. In our system we used a simple array-based implementation for mapping tables.
There are a couple of options for implementing job queues. Recall that recovery-consciousness is built on top of the QoS or priority based scheduling. We could use multiple QoS based job queues (for example, high, medium and low priority queues) for each pool or for each group. In our first prototype, we chose the latter option and implemented multiple QoS based job queues for each recovery group for a number of reasons. Firstly, this choice easily fits into the scenario where a single recovery group is assigned to multiple resource pools. Secondly, it offers greater flexibility to modify mappings at runtime. Finally, reactive scheduling (i.e., suspending dispatch of tasks belonging to a group undergoing localized recovery) can be implemented more elegantly as the resource scheduler can simply skip the job queues for the recovering group. Enqueue and dequeue operations on each queue are protected by a lock. An additional advantage of a mapping implemented using multiple independent queues is that it reduces the degree of contention for queue locks.
The second class of algorithms are partially dynamic and allow the recovery-conscious scheduler to react (in a constrained fashion though) to sudden spikes or bursty workload of a recovery group.
The main drawback of static RCS is that it results in poor utilization of resources due to the strictly fixed mapping. Partial dynamic RCS attempts to alleviate this problem by using a relatively more flexible mapping of resources to recovery groups, allowing groups to utilize spare resources. Partially dynamic RCS algorithms begin with a static mapping. However, when the utilization is low, the system switches to an alternative mapping that redistributes resources across recovery groups.
For example, with the static mapping of Table 2 with changing distribution of workloads, resources allocated to recovery groups γ3 and γ4 may be under utilized while groups γ1 and γ2 may be swamped with work. Under these circumstances, the system switches to an alternative mapping shown in Table 3. Now groups γ1 and γ2 can utilize spare resources across the system even if this may mean potentially violating their recoverability constraints specified in Table 1. Note that γ3 and γ4 still obey their recoverability constraints. In summary, partially dynamic mappings allows the flexibility of selectively violating the recoverability constraints when there are spare resources to be utilized, whereas static mappings strictly obey recoverability constraints.
Implementation Considerations: There are two implementation considerations that are specific to the partially dynamic scheduling schemes: (1) the mechanism to switch between initial schedule and an alternative schedule, and (2) the mapping of recovery group tasks to the shared resource pools.
We use a lease expiry methodology to flexibly switch between alternative mappings. Note that the pool schedulers switch to the alternative mapping based on the resource utilization of the current pool. With the partially dynamic scheme, the alternative mappings are acquired under a lease, which upon expiry causes the scheduler to switch back to the original schedule. The lease-timer is set based on observed component workload trends (such as duration of a burst or spike) and the cost of switching. For example, since our implementation had a very low cost of switching between mappings, we set the lease-timer to a single dispatch cycle. Figure 8 shows the pseudo-code for a partial dynamic scheduling scheme using a lease expiry methodology. For the sake of simplicity we do not show the tracking method (round-robin) used to avoid starvation in the scheduler.
Recall from the implementation considerations for the static mapping case that we chose to implement job queues on a per recovery group basis. This allowed for easy switching between the current and alternative mapping which only involves consulting a different mapping table. Task enqueue operations are unaffected by the switching between mappings.
This class of algorithms aim at maximizing utilization of resources at the cost of selectively violating the recoverability constraints. Note that all recovery-conscious algorithms are still designed to perform reactive scheduling, i.e., suspend the dispatching of tasks whose group is currently undergoing localized recovery. The aspect that differentiates the various mapping schemes is the proactive handling of tasks to improve system availability. The dynamic scheme can be thought of as trying to use load balancing among recovery groups in order to achieve both recovery isolation and good resource utilization.
Implementation Considerations: A key implementation consideration specific to dynamic RCS is the problem of keeping track of the number of outstanding tasks belonging to each recovery group. We maintain this information in a per-processor data structure that keeps track of the current job.
Recall that implementing job queues on the per recovery-group basis helps us implement dynamic mappings efficiently and flexibly. One of the critical optimizations for dynamic RCS algorithms involves understanding and mitigating the scheduling overhead imposed by the dynamic dequeuing process. In on going work we are conducting experiments with different setups to characterize this overhead. However our results in this paper show that even with the additional scheduling cost dynamic RCS schemes perform very well both under good-path and bad-path conditions.
The embedded storage controller software is similar to the model presented in Section 2. The software is also highly-reliable with provisions for quick recovery (under ~ 6 seconds) at the system-level. The system has a number of interacting components which dispatch a large number of short running tasks. For the experiments in this paper based on programmer explicit recovery dependency specifications, we identified 16 recovery groups which roughly correspond to functional components such as cache manager, device manager, SCSI command processor etc. However, some recovery groups may perform no work in certain workloads possibly due to features being turned off. We chose a pool size of 1 CPU which resulted in 8 pools of equal size. The system already implements high, medium and low priority job queues. Our recovery-conscious scheduling implementation therefore uses three priority based queues per recovery group. For the partially dynamic case, based on the workload we have identified two candidates for strict isolation - groups 4 and 5. For the static mapping case each recovery group is mapped to resource pools proportional to its ratio of the total task workload.
Figure 11 shows the good-path throughput for the performance-oriented and recovery-conscious scheduling algorithms. The average throughput for the dynamic RCS case with 16 groups (105 KIOps) and 4 groups (106 KIOps) was close to that for the performance-oriented scheduler (107 KIOps). On the other hand, with partially dynamic RCS, the system throughput drops by nearly 34% (~ 69.9 KIOps), and with static RCS by nearly 58% (~ 44.6 KIOps) compared to performance oriented scheduling.
Figure 12 compares the response time with different RCS schemes and performance-oriented scheduling. Again, the average response-time for the dynamic RCS case with 16 recovery groups (13.5 ms) and 4 recovery groups (13.6 ms) is close to the performance-oriented case (13.3 ms). However, with the partially-dynamic RCS scheduling, the response time increases by nearly 63% (21.7 ms) and by 156% with static RCS (34.1ms).
Both the throughput and response time numbers can be explained using the next chart, Figure 13. The radar plot shows the relationship between throughput, response time and cpu-utilization for each of the cases. As the figure shows, the cpu utilization has dropped by about 19% with partially dynamic RCS and by 63% with static RCS. The reduction in cpu utilization eventually translates to reduced throughput and increased response time in a cpu-bound workload intensive environment. These numbers seem to indicate that in such an environment schemes that reduce the utilization can result in significant degradation of the overall performance. However note that the normal operating range of many customers may be only around 6-7000 IOps . If that be the case, then partially-dynamic schemes can more than meet the system requirement even while ensuring some recovery isolation.
In the case of recovery-conscious scheduling algorithms, reactive scheduling kicks in when any group is undergoing recovery. Under those circumstances, tasks belonging to that recovery group already under execution are allowed to finish, but further dispatch from that group is suspended until recovery completes.
Figure 14 shows the average system throughput with fault injection. The average throughput using only performance oriented scheduling (87.8 KIOps) drops by nearly 16.3% when compared to dynamic RCS (105 KIOps) that also uses reactive policies. On the other hand dynamic RCS continues to deliver the same throughput as under normal conditions. Note that this is still not the worst case for performance oriented scheduling. In the worst case, all resources may be held up by the recovering tasks resulting in actual service outage and the problem would only worsen with increasing localized recovery time and system size.
The figure also compares proactive and reactive policies in dynamic RCS. The results show that with only proactive scheduling we are able to sustain a throughput (104 KIOps) which is just ~ 1% less than that using both proactive and reactive policies (105 KIOps).
The graph also compares partially dynamic RCS (69.9 KIOps) and static RCS (40.4 KIOps). While these schemes are able to sustain almost the same throughput as they do under good path, overall, the performance of these schemes results in 20% and 54% drop in throughput respectively compared to performance oriented scheduling.
We performed experiments with other configurations of dynamic, partially dynamic and static schemes and using other workloads too. However due to space constraints we only present key findings from those experiments. In particular we used a disk-bound internal workload (and hence low cpu utilization of about ~ 25%) to study the effect of our scheduling algorithms under a sparse workload. We used the number of task dispatches as a metric of scheduler efficiency. The fault injection mechanism was similar to the cache-standard workload, however due to the workload being sparse, we introduced an overhead of only 0.3% to the aggregate execution time of the faulty recovery group. Our results showed that dynamic RCS was able to achieve as many dispatches as performance oriented scheduling under good path operation and increase the number of dispatches by 0.7% under bad-path execution. With partial dynamic RCS dispatches dropped by 20% during good path operation and by only 3.9% during bad path operation compared to performance oriented scheduling. The same static mapping used in the cache standard workload when run in this new environment resulted in the system not coming up. While this may be due to setup issues, it is also likely that insufficient resources were available to the platform tasks during start-up. We are investigating further on a more appropriate static mapping for this environment.
Dynamic RCS can match good path performance of performance oriented scheduling and at the same time significantly improve performance under localized recovery. Even for the small 5% recovery overhead introduced by us, we could witness a 16.3% improvement in throughput and a 22.9% improvement in response time with dynamic RCS. Moreover, the qualitative benefits of RCS in enhancing availability and ensuring that localized recovery is scalable with system size makes it an interesting possibility as systems are moving towards more parallel architectures. Our experiments with various scheduling schemes (some not reported in this paper) have given us some insights into the overhead costs such as lock spin times imposed by RCS algorithms. In ongoing work we are continuing to characterize and investigate further optimizations to RCS schemes.
Our results also seem to indicate that for small localized recovery time and system sizes, proactive policies i.e. mapping resource pools to recovery groups, can deliver the advantage of recovery-consciousness. However as system size increases or localized recovery time increases, we believe that the actual benefits of reactive policies such as suspending dispatch from groups undergoing recovery may become more pronounced. In ongoing research we are experimenting with larger setups and longer localized recovery times.
Static and partial dynamic RCS schemes are limited by their poor resource utilization in workload intensive environments. Hence we do not recommend these schemes in an environment where the system is expected to run at maximum throughput. However, the tighter qualitative control that these schemes offer may make them, especially partially dynamic RCS, more desirable in less intensive environments where there is a possibility to over-provision resources, or when the workload is very well understood. Besides in environments where it `pays' to isolate some components of the system from the rest such mappings may be useful. We are continuing research on optimizing these algorithms and understanding properties that would prescribe the use of such static or partially dynamic schemes.
In general our recovery conscious approaches are complementary to the above techniques. However we are faced with several unique challenges in the context of embedded storage software. First, the software being legacy code rules out re-architecting the system. Second, the tight coupling between components makes both micro-reboots and periodic rejuvenation tricky. Rx  demonstrates an interesting approach to recovery by retrying operations in a modified environment but it requires checkpointing of the system state in order to allow `rollbacks'. However given the high volume of requests (tasks) experienced by the embedded storage controller and their complex operational semantics, such a solution may not be feasible in this setup.
Failure-oblivious computing  introduces a novel method to handle failures - by ignoring them and returning possibly arbitrary values. This technique may be applicable to systems like search engines where a few missing results may go unnoticed, but is not an option in storage controllers.
The idea of localized recovery has been exercised by many. Transactional recovery using checkpointing/logging methods is a classic topic in DBMSs  and is a successful implementation of fine-grained recovery. In fact application-specific recovery mechanisms such as recovery blocks , and exception handling  are used in almost every software system. However, very few have made an effort on understanding the implications of localized recovery on system availability and performance in a multi-core environment where interacting tasks are executing concurrently. Likewise, the idea of recovery-conscious scheduling is to raise the awareness about localized recovery in the resource scheduling algorithms to ensure that the benefits of localized recovery actually percolate to the level of system availability and performance visible to the user. Although vast amounts of prior work have been dedicated to resource scheduling, to the best of our knowledge, such work has mainly focused on performance [25,11,12,8,4]. Also much work in the virtualization context has been focused on improving system reliability  by isolating VMs from failures at other VMs. In contrast, our development focuses more on improving system availability by distributing resources within an embedded storage software system by identifying fine-grained recovery scopes. Compared to earlier work on improving storage system availability at the RAID level , we are concerned with the embedded storage software reliability. These techniques are at different levels of the storage system and are complementary.
In order to adopt RCS for large software systems, a significant challenge is to identify efficient recovery scopes. In ongoing work we are working on developing more generic guidelines that would assist in identifying fine-grained recovery scopes. Even with pluggable mechanisms like RCS it is necessary to emphasize that high-availability should still be a design concern and not an after-thought. We hope our framework would encourage developers to incorporate additional error handling and anticipate more error scenarios and that our scheduling schemes would aid in scaling efficient error handling with system size.
This work is partially supported by an IBM PhD scholarship and an IBM Almaden Storage Systems internship for the first author, and the NSF CISE IIS and CSR grants, an IBM SUR grant, as well as an IBM faculty award for the authors from Georgia Tech.
This document was generated using the LaTeX2HTML translator Version 2002-2-1 (1.70)
Copyright © 1993, 1994, 1995, 1996,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.