M. Devarakonda and A. Mukherjee. Issues in implementation of cache-affinity scheduling. In Proc. of Winter USENIX Conference, pages 345--357, Jan. 1992.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....assignment by different scheduling schemes. Since applications 2 and 3 behave similarly, only one set of figures is shown. 5 Related Work Cache affinity scheduling, which uses of locality information for the scheduling decision has been used mostly in shared memory multiprocessors [16] [7], 12] 15] The focus in this domain is to schedule the same process thread on processors that can reuse previously established cache state. While this is similar to the network processor environment, it does not consider the reuse of instruction cache state for different threads that use the ....

M. Devarakonda and A. Mukherjee. Issues in implementation of cache-affinity scheduling. In Proc. of Winter USENIX Conference, pages 345--357, Jan. 1992.

....SGI Challenge XL) packet execution times can vary by as much as a factor of four, depending on the state of the processor cache. This suggests that affinity based scheduling techniques (which have been explored in the context of general parallel programs (i.e. non network application processing) [15, 28, 46, 68, 76]) present a promising research opportunity in parallel networking. We have recently completed an extensive study in this area [61, 63, 62] the results are summarized in Section 3 and several extensions proposed. We also seek to explore the design of parallel networking for communication intensive ....

....sharing a given context record. Via simulation and analytic techniques, the authors consider appropriate processor group size for minimized delay in servicing an individual signaling request. 6. 2 Affinity based scheduling Four studies consider affinity based scheduling at the process level [15, 28, 68, 76], while one considers scheduling at a finer granularity [46] In general, the earlier studies [15, 28, 76] did not find affinity scheduling to be of significant benefit to common applications. In these cases, performance gain was precluded by the fact that the time required for the scheduled task ....

[Article contains additional citation context not shown here]

Murthy Devarakonda and Arup Mukherjee. Issues in implementation of cache-affinity scheduling. In Proceedings of the Winter 1992 USENIX Conference, pages 345--357, San Francicso, CA, January 1992.

....is assigned a time slice that the processor should devote to its task. The scheduled task is suspended when the time slice is over, or as a result of system call such barrier, semaphores and locks. As the task is suspended, it returns to the waiting queue to be rescheduled. Several works ([Dava92] and [Vasw91] suggested to employ cache affinity; i.e. to give higher priority to rescheduling the tasks on its recent processor. Unfortunately the overhead to maintain such history information was found to be too high in respect to the benefit that can be obtained. Moreover, we believe that ....

M. Devarakonda and A. Mukherjee. Issues in Implementation of Cache-Affinity Scheduling. USENIX, 1992.

....policies. The hardware platform serving as the basis for the experimental component of our study is an 8 processor SGI Challenge XL running the IRIX 5.2 operating system. Protocols are implemented using the x kernel framework in user space [8, 14] Previous work on affinity based scheduling [3, 5, 22, 26] has not established a consensus on its efficacy, even reaching seemingly conflicting conclusions (e.g. 26] vs. 22] None of the work has found affinity based scheduling techniques to be of significant benefit to common applications. For affinity scheduling to be effective, the enabled ....

....in section 4.4 coupled with additional analytic results from Singh, Stone and Thiebaut [20] However, Squillante and Lazowska do not identify specific applications which stand to benefit from affinity scheduling. We have shown that parallelized protocol processing is one such application. In [3], Devarakonda and Mukherjee explore implementation issues in affinity scheduling, both in kernel and within a user level thread scheduler, on an 8 processor Encore Multimax running Mach 2.5. A schedulable task is defined to have affinity strictly for the processor it most recently visited. They ....

[Article contains additional citation context not shown here]

Murthy Devarakonda and Arup Mukherjee. Issues in implementation of cache-affinity scheduling. In Proceedings of the Winter 1992 USENIX Conference, pages 345--357, San Francicso, CA, January 1992.

....boils down to the question of data that may reside in the PE s cache. The scheduling policy that tries to schedule threads on the same PE on which they ran most recently, under the assumption that this particular PE might still have some relevant data in its cache, is called affinity scheduling [598, 157, 38, 544, 55]. Affinity scheduling becomes more important on machines where the cost of remote access is higher [106] However, the performance improvement is typically small, as in most cases not much is left in the cache after a number of other applications have been scheduled [249, 576, 598] It has ....

....caches. A number of studies have demonstrated the adverse effect of context switching on cache performance [408, 249, 598] Affinity scheduling, where threads are scheduled back onto previously used PEs so as to benefit from data that may still reside in their caches, tries to counter this effect [598, 157, 38, 544, 576, 106]. It has even been proposed that affinity hints be included in the programming language [107] Note that the adverse effects caused by the operating system are related in both cases to the preemption of running threads. Thus memory usage considerations lead to a preference for non preemptive ....

M. Devarakonda and A. Mukherjee, "Issues in implementation of cache-affinity scheduling ". In Proc. Winter USENIX Technical Conf., pp. 345--357, Jan 1992.

....boils down to the question of data that may reside in the PE s cache. The scheduling policy that tries to schedule threads on the same PE on which they ran most recently, under the assumption that this particular PE might still have some relevant data in its cache, is called affinity scheduling [348, 86, 25, 328, 33]. Affinity scheduling becomes more important on machines where the cost of remote access is higher [57] However, the performance improvement is typically small, as in most cases not much is left in the cache after a number of other applications have been scheduled [152, 339, 348] It has ....

.... This effect becomes more noticeable the smaller the memory, and is therefore especially relevant for caches [249, 152, 348] Affinity scheduling, where threads are scheduled back onto previously used PEs so as to benefit from data that may still reside in their caches, tries to counter this effect [348, 86, 25, 328, 339, 57]. Effect on communication locality Communication is an important aspect of parallel applications. In many systems, communication costs depend on distance. However, this dependency has been decreasing in new architectures, due to wormhole routing and the use of lower dimensionality networks. ....

M. Devarakonda and A. Mukherjee, "Issues in implementation of cache-affinity scheduling ". In Proc. Winter USENIX Technical Conf., pp. 345--357, Jan 1992.

....if relevant data already resides in the processor s cache. This type of processor affinity is particularly interesting, because it may diminish with time, and because it is dependent on fairly complex characteristics of the workload. There are a significant number of references in the literature [19] [24] 26] 41] 43] 53] 54] 57] 60] 63] that deal with this type of processor affinity; this work is the subject of the first two sections of this chapter. In the remaining section, two new affinity scheduling algorithms are proposed. 3.1 Principles and Algorithms of Affinity Scheduling A ....

....the basic affinity scheduling algorithm) varied from 0 7 (depending upon the load and workload characteristics) and must be traded off with the disadvantages of a more complex implementation. A study of cache affinity scheduling with the help of synthetic workloads by Devarakonda and Mukherjee [19] emphasized the point that previous studies showing negative results for affinity scheduling were with respect to kernel level affinity scheduling (coarse grained kernel threads) and that the kernel may not be the appropriate place to implement affinity scheduling. This study also pointed out ....

M. Devarakonda, A. Mukherjee, "Issues in the Implementation of Cache-Affinity Scheduling", Proceedings of the Winter 1992 USENIX Conference, January 1992, pp. 345-357.

....each time they arrive at the head of the queue. As a result, any state that may be left in a processor s cache is lost. It has been suggested that this effect can be reduced by using affinity scheduling, where an effort is made to re schedule the process on the same processor as used last time [49,10]. However, it is not clear to what degree data indeed remains in the cache, and in any case, affinity scheduling is largely equivalent to just using longer time quanta [51] The third issue, lack of coordinated scheduling, may cause problems for applications where the processes interact with each ....

M. Devarakonda and A. Mukherjee, "Issues in implementation of cache-affinity scheduling ". In Proc. Winter USENIX Technical Conf., pp. 345--357, Jan 1992.

....to be found in that processor s cache, thus avoiding accesses to the slower main memory and resulting in faster execution times. Previous studies have explored the benefits of affinitybased scheduling in the context of general parallel programs (i.e. non network related application processing) [4, 6, 12, 24, 27]. In this paper, we explore affinitybased scheduling of parallel networking, an area of research which has recently generated considerable interest (e.g. 3, 11, 13, 19, 21] In general, for affinity based scheduling to be effective the time between rescheduling of the affinity managed resource ....

....performs better from the perspective of the host. 6 Related Work In this section, we briefly review prior work in processorcache affinity scheduling. No previous study has examined the technique in the context of parallel network processing. Four studies consider scheduling at the process level [4, 6, 24, 27], while one considers the finer granularity of loop scheduling [12] Vaswani and Zajorian [27] show experimentally that affinity based scheduling within kernel level processor space sharing scheduling policies provides little benefit. Their workload is a mix of three types of parallel applications ....

[Article contains additional citation context not shown here]

M. Devarakonda and A. Mukherjee. "Issues in Implementation of Cache-Affinity Scheduling". Proceedings of the Winter USENIX Conference, p. 345-357, Jan. 1992.

....which improves the caching behavior and available packetlevel concurrency under connection level parallelism, and find performance improves dramatically. 1 Introduction Processor cache affinity scheduling is of growing interest as processor speeds continue to increase faster than memory speeds [6, 9, 21, 30, 37, 39]. On modern shared memory machines, the time required to access an uncached memory location is typically much larger than the time to access one cached locally. For example, on the Silicon Graphics Challenge XL multiprocessor, the experimental platform used in our research, a memory reference can ....

....from 115 s to 130 s to 210 s in Figure 13, whereas it remains relatively stable across values of B under PLP (Figure 10) 9 Related Work In this section, we briefly review other work in processorcache affinity scheduling. Four studies consider affinitybased scheduling at the process level [6, 9, 37, 39], while one considers the finer granularity of loop scheduling [21] Vaswani and Zajorian [39] show experimentally that affinity based scheduling within kernel level processor spacesharing scheduling policies provides little benefit. Their workload is a mix of three types of parallel ....

[Article contains additional citation context not shown here]

Murthy Devarakonda and Arup Mukherjee. Issues in implementation of cache-affinity scheduling. In Proc. Winter 1992 USENIX Conference, pages 345--357, San Francicso, CA, January 1992.

....that processor s cache, thus avoiding accesses to the slower main memory and resulting in faster execution times. A number of previous studies of have explored the benefits of affinity based scheduling in the context of general parallel programs (i.e. non network related application processing) [5, 8, 19, 30, 34], with somewhat conflicting results. Althougha primarily analytic study has demonstrated significant potential benefits of the technique [30] most implementation studies of process level affinity scheduling have found only a marginal improvement for common applications [5, 8, 34] This is because ....

....[5, 8, 19, 30, 34] with somewhat conflicting results. Althougha primarily analytic study has demonstrated significant potential benefits of the technique [30] most implementation studies of process level affinity scheduling have found only a marginal improvement for common applications [5, 8, 34]. This is because the time between rescheduling of the affinity managed entity (in these cases, the process) is much larger than the time required to entirely reload the referenced memory locations into the cache. In this paper, we explore the benefits of affinity scheduling of parallel ....

[Article contains additional citation context not shown here]

Murthy Devarakonda and Arup Mukherjee. Issues in implementation of cache-affinity scheduling. In Proceedings of the Winter 1992 USENIX Conference, pages 345--357, San Francicso, CA, January 1992.

....node from a set of threads that have run on that node recently. This strategy is motivated by the assumption that the thread will continue to access the data it has touched in the recent past and that the data will still be in the cache at that processor. These techniques have met with some success[8, 13, 26] in the context of heavyweight Unix processes using coarse grain styles of multiprogramming. In contrast, a lightweight thread created to perform a single task does not accumulate much history. It may require only one scheduling decision over its entire lifetime. For such threads, few ....

....is done by changing isolated 1 s to 0 s, and processing stops when no more changes are possible. Each phase is divided into two subphases, each of which uses slightly different definition of isolated . In its coarse grain version, this application was identified by Devarakonda and Mukherjee [8] as being ideally suited to cache affinity scheduling because threads synchronize four times per phase. A medium grain version can also benefit when threads are rescheduled between sub phases, but creating independent threads for each phase limits the potential of this strategy. We measured five ....

[Article contains additional citation context not shown here]

M. Devarakonda and A. Mukherjee. "Issues in the Implementation of CacheAffinity Scheduling". In Proceedings of Winter '92 USENIX Conference, 1992.

....Cache affinity scheduling [22, 12, 24] attempts to reschedule a process on the processor it last used, thereby reducing process migration and cache interference. The effectiveness of cache affinity has been evaluated through analytical studies [22, 21] simulation [12] and actual implementations [24, 8, 27]. Experiments on bus based multiprocessors [24] however, show that realistic applications achieve only moderate gains (less than 10 ) from affinity scheduling. We will show that the gains can be much larger on CC NUMA machines. Gang Scheduling: For parallel applications, Ousterhout [19] proposed ....

M. Devarakonda and A. Mukherjee. Issues in implementation of cache-affinity scheduling. In Proceedings Winter 1992 USENIX Conference, pages 345--357, January 1992.

....short 16 ms time quantum. Simulations driven by application only traces of real multiprocessor applications by Gupta et al. 4] with a short time quantum (10 ms) point to small benefits of affinity. However, I O and other load variations were not considered. A study by Devarakonda and Mukherjee [2] evaluated the performance of a real implementation of cache affinity on a multiprocessor, but used synthetic workloads. Their results suggest that implementation issues and workload choice can have a large impact on the measured performance of affinity scheduling, but the synthetic nature of the ....

M. Devarakonda and A. Mukherjee. Issues in Implementation of Cache-Affinity Scheduling. In Proceedings Winter 1992 USENIX Conference, pages 345--357, January 1992.

....by NSF under grant NCR 9206908 and by ARPA under ESD AVS contract F 19628 92 C 0089. The authors can be contacted at [salehi,kurose,towsley] cs.umass.edu. 1 Introduction Processor cache affinity scheduling is of growing interest as processor speeds continue to increase faster than memory speeds [7, 10, 22, 32, 38, 40]. On modern shared memory machines, the time to access an uncached memory location is typically much larger than when accessing one cached locally. For example, on the SGI Challenge XL multiprocessor (the experimental platform used in our research) a memory reference can be serviced by the ....

....parallelism in the host operating system both within and among connections can both increase the bandwidth and decrease the latency of multiprocessor communication. While previous studies have explored the benefits of affinity based scheduling of non network related application processing [7, 10, 22, 40], our work [31, 32, 33] is the first to apply the technique to operating system network processing. In our experimental environment (consisting of a parallelized x kernel [13, 26] running in user space on an 8 processor MIPS R4400 based SGI Challenge XL) packet execution times can vary by as much ....

Murthy Devarakonda and Arup Mukherjee. Issues in implementation of cache-affinity scheduling. In Proceedings of the Winter 1992 USENIX Conference, pages 345--357, San Francicso, CA, January 1992.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC