Compiler Assisted Load Balancing on Large Clusters

Deodhar, Vinit; Parikh, Hrushit; Gavrilovska, Ada; Pande, Santosh

doi:10.1109/pact.2015.40

Cited by 5 publications

(2 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…[23] proposes a compiler-driven approach to reduce power consumption by inserting statements to shut down functional units through a profile-driven approach. [6] is also a compiler-based technique for load-balancing and proposes early notifications before loops, considers only floating point and mem-op instructions within loops for predicting resource usage, and places more stress on inter-procedural placement of notifications in cluster nodes. The QoS-Compile [25] proposes a static compilation framework to deal with the cache contention problem of co-locating low priority and high priority processes.…”

Section: Related Workmentioning

confidence: 99%

Compiler-Guided Throughput Scheduling for Many-core Machines

Mururu¹,

Khan²,

Chatterjee³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Modern ARM-based servers such as ThunderX and Thun-derX2 offer a tremendous amount of parallelism by providing dozens or even hundreds of processors. However, exploiting these computing resources for reuse-heavy, data dependent workloads is a big challenge because of shared cache resources. In particular, schedulers have to conservatively co-locate processes to avoid cache conflicts since miss penalties are detrimental and conservative co-location decisions lead to lower resource utilization.To address these challenges, in this paper we explore the utility of predictive analysis of applications' execution to dynamically forecast resource-heavy workload regions, and to improve the efficiency of resource management through the use of new proactive methods. Our approach relies on the compiler to insert "beacons" in the application at strategic program points to periodically produce and/or update the attributes of anticipated resource-intense program region(s). The compiler classifies loops in programs based on predictability of their execution time and inserts different types of beacons at their entry/exit points. The precision of the information carried by beacons varies as per the analyzability of the loops, and the scheduler uses performance counters to fine tune co-location decisions. The information produced by beacons in multiple processes is aggregated and analyzed by the proactive scheduler to respond to the anticipated workload requirements. For throughput environments, we develop a framework that demonstrates high-quality predictions and improvements in throughput over CFS by 1.4x on an average and up to 4.7x on ThunderX and 1.9x on an average and up to 5.2x on ThunderX2 servers on consolidated workloads across 45 benchmarks.

show abstract

Section: Related Workmentioning

confidence: 99%

Compiler-Guided Throughput Scheduling for Many-core Machines

Mururu¹,

Khan²,

Chatterjee³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…An important aspect of the matchmaking framework is its flexibility to allow different spatial or temporal metrics to be incorporated in the scheduler. For instance, in one design-matchmakers can be strategically selected based on cluster topology which leads More interestingly, for workloads where future load supply or demand can be predicted at run-time (as shown in our prior work [2])temporally proximate producer/consumer requests can be sent to a designated matchmaker-i.e producer/consumer requests anticipated for a future time window W1 are sent to a particular matchmaker and requests anticipated for window W2 sent to a different matchmaker. In this way, matchmakers are responsible for anticipated load balancing for specific, smaller temporal windows, thus resulting in ability to scale further in a distributed manner.…”

Section: Design Considerationsmentioning

confidence: 99%

Efficient distributed workstealing via matchmaking

et al. 2016

Self Cite

View full text Add to dashboard Cite

Many classes of high-performance applications and combinatorial problems exhibit large degree of runtime load variability. One approach to achieving balanced resource use is to over decompose the problem on fine-grained tasks that are then dynamically balanced using approaches such as workstealing. Existing work stealing techniques for such irregular applications, running on large clusters, exhibit high overheads due to potential untimely interruption of busy nodes, excessive communication messages and delays experienced by idle nodes in finding work due to repeated failed steals. We contend that the fundamental problem of distributed work-stealing is of rapidly bringing together work producers and consumers. In response, we develop an algorithm that performs timely, lightweight and highly efficient matchmaking between work producers and consumers which results in accurate load balance. Experimental evaluations show that our scheduler is able to outperform other distributed work stealing schedulers, and to achieve scale beyond what is possible with current approaches.

show abstract