Enabling Fair Pricing on High Performance Computer Systems with Node Sharing

Abstract

Abstract. Co-location, where multiple jobs share compute nodes in large-scale HPC systems, has been shown to increase aggregate throughput and energy efficiency by 10-20%. However, system operators disallow co-location due to fair-pricing concerns, i.e., a pricing mechanism that considers performance interference from co-running jobs. In the current pricing model, application execution time determines the price, which results in unfair prices paid by the minority of users whose jobs suffer from co-location. This paper presents POPPA, a runtime system that enables fair pricing by delivering precise online interference detection and facilitates the adoption of supercomputers with co-locations. POPPA leverages a novel shutter mechanism -a cyclic, fine-grained interference sampling mechanism to accurately deduce the interference between co-runners -to provide unbiased pricing of jobs that share nodes. POPPA is able to quantify inter-application interference within 4% mean absolute error on a variety of co-located benchmark and real scientific workloads.

Caliper

“…Such slowdown estimates could enable resource allocation of shared resources to each application in a slowdown-aware manner motivated towards providing strong Quality-of-Service (QoS) guarantees. Also, in Infrastructure-as-a-Service (IaaS) clouds, such a mechanism could be used to bill its customers appropriately based on the amount of slowdown that their applications have been subjected to by the co-running applications [16,58].…”

“…There have been many efforts that try to estimate slowdown of applications at runtime [16,17,23,25,40,42,49,58,61,68]. Prior software approaches [16,25,61,68] utilize an online runtime system that periodically pauses all the applications except one for a short time, thus allowing the running application to monopolize the computing resources on the system during those pause periods. The performance of the running application during such pause periods is used to determine slowdown.…”

“…(1) Low accuracy: The most recent state-of-the-art technique addressing this problem [16] neglects the notion of application phases and pauses co-running applications periodically at millisecond granularity. This methodology shows estimation errors of up to 40%, leaving significant room for improvement in accuracy.…”

“…to 2% degradation in performance to support dynamic monitoring approaches in production [55]. However, the execution time overhead of the state-of-the-art software interference estimation technique can be as significant as 12% [16]. (3) Non-reliable (or less scalable): The accuracy and the overhead of prior approaches [16,23,25,42,58] deteriorate as the number of co-running applications increases.…”

“…However, the execution time overhead of the state-of-the-art software interference estimation technique can be as significant as 12% [16]. (3) Non-reliable (or less scalable): The accuracy and the overhead of prior approaches [16,23,25,42,58] deteriorate as the number of co-running applications increases. As the number of cores on modern servers keeps increasing, deploying a technique that inadequately supports current and future levels of multi-tenancy would not be a preferred choice.…”