Adaptive insertion policies for managing shared caches

Jaleel, Aamer; Hasenplaugh, William; Qureshi, Moinuddin K.; Sébot, Julien; Steely, Simon C.; Emer, Joel

doi:10.1145/1454115.1454145

Cited by 255 publications

(173 citation statements)

References 18 publications

Supporting

Mentioning

173

Contrasting

Order By: Relevance

“…Recent work has noted the poor scalability of having separate monitors for each core and methods have been proposed including In-Cache Estimation Monitors [11] and set-dueling [12] to eliminate the need for separate monitors. A number of sets in the cache are dedicated to a particular core from which the monitored statistics can be gathered.…”

Section: Related Workmentioning

confidence: 99%

Set Utilization Based Dynamic Shared Cache Partitioning

Deayton

Chung

2011

2011 IEEE 17th International Conference on Parallel and Distributed Systems

View full text Add to dashboard Cite

Abstract-As the number of processors sharing a cache increases, conflict misses due to interference amongst competing processes have an increasing impact on the individual performance of processes. Cache partitioning is a method of allocating a cache between concurrently executing processes in order to counteract the effects of inter-process conflicts. However, cache partitioning methods commonly divide a shared cache into private partitions dedicated to a single processor, which can lead to underutilized portions of the cache when set accesses are non-uniform.Our proposed method compliments these cache partitioning algorithms by creating an additional shared partition able to be shared amongst all processors. Underutilized areas of the cache are identified by a monitoring circuit and used for the shared partition. Detection of underutilization is based on the number of unique set accesses for a given allocated way. For a 16-way set associative cache, the implementation of our method requires 64 bytes of storage overhead per core in addition to that needed for the method that determines the sizes of the private partitions. For the tested system, our method is able to improve performance over the traditional LRU policy for a number of selected benchmark sets by an average of 1.4% and up to 13.3% for a two core system and an average of 1.4% and up to 7.8% for a four core system, and is able to improve the performance of a conventional cache partitioning method (Utility-Based Cache Partitioning) by an average of 0.1% and up to 0.5% for both a two and four core systems.

show abstract

Section: Related Workmentioning

confidence: 99%

Set Utilization Based Dynamic Shared Cache Partitioning

Deayton

Chung

2011

2011 IEEE 17th International Conference on Parallel and Distributed Systems

View full text Add to dashboard Cite

show abstract

“…Recent work has shown that level two cache miss rates can be improved by preventing or limiting the amount of data retained for applications with streaming data accesses or working sets larger than the level two cache capacity [16]. For shared caches, different data retention policies can be applied to different processes in order to maximize the use of the shared capacity [17]. Other approaches like cache decay can target the same types of data in order to reduce the used capacity in a cache by turning off lines that have remained idle for long periods of time [18].…”

Section: A Cache Design and Managementmentioning

confidence: 99%

“…As the different types of data exhibit different characteristics with respect to idle frequency and idle duration, it may also be beneficial to tailor these techniques to specific types of data as suggested as future work in [17]. For example, data in ScalParC that is shared by many threads goes idle frequently and for long periods of time, so it might be worthwhile to prevent it from being cached.…”

Section: Data Reusementioning

confidence: 99%

Understanding the applicability of CMP performance optimizations on data mining applications

Jibaja

Shaw

2009

2009 IEEE International Symposium on Workload Characterization (IISWC)

View full text Add to dashboard Cite

Abstract-A major challenge to the creation of chip multiprocessors is designing the on-chip memory and communication resources to efficiently support parallel workloads. A variety of cache organizations, data management techniques, and hardware optimizations that take advantage of specific data characteristics have been developed to improve application performance. The success of these approaches depends on applications exhibiting the presumed data characteristics.Data mining applications are a growing class of applications that discover patterns in large sets of collected data. Because these applications tend to be highly parallelizable, they represent an important workload for chip multiprocessors. However, the memory intensive nature of these applications means that they will stress these chips' memory and communication resources.In this paper, we examine the data usage characteristics of a set of parallel data mining applications to determine the applicability of existing chip multiprocessor approaches to these applications. We show diversity of characteristics across and within these applications, making some techniques more applicable than others. We also discuss software approaches that could be used to either provide information to the hardware or assist the hardware in dynamically discovering data characteristics needed for the deployment of these techniques.

show abstract

“…Shared caches in contemporary multicores have repeatedly been shown to be critical resources for performance [15,23,28,8,17]. A significant amount of research has investigated the impact of cache sharing on application performance [23,30,12,11].…”

Section: Introductionmentioning

confidence: 99%

Modeling performance variation due to cache sharing

Sandberg

Sembrant

Hägersten

et al. 2013

2013 IEEE 19th International Symposium on High Performance Computer Architecture (HPCA)

View full text Add to dashboard Cite

Shared cache contention can cause significant variability in the performance of co-running applications from run to run. This variability arises from different overlappings of the applications' phases, which can be the result of offsets in application start times or other delays in the system. Understanding this variability is important for generating an accurate view of the expected impact of cache contention. However, variability effects are typically ignored due to the high overhead of modeling or simulating the many executions needed to expose them. This paper introduces a method for efficiently investigating the performance variability due to cache contention. Our method relies on input data captured from native execution of applications running in isolation and a fast, phaseaware, cache sharing performance model. This allows us to assess the performance interactions and bandwidth demands of co-running applications by quickly evaluating hundreds of overlappings.We evaluate our method on a contemporary multicore machine and show that performance and bandwidth demands can vary significantly across runs of the same set of co-running applications. We show that our method can predict application slowdown with an average relative error of 0.41% (maximum 1.8%) as well as bandwidth consumption. Using our method, we can estimate an application pair's performance variation 213× faster, on average, than native execution.

show abstract

Adaptive insertion policies for managing shared caches

Cited by 255 publications

References 18 publications

Set Utilization Based Dynamic Shared Cache Partitioning

Set Utilization Based Dynamic Shared Cache Partitioning

Understanding the applicability of CMP performance optimizations on data mining applications

Modeling performance variation due to cache sharing

Contact Info

Product

Resources

About