Heterogeneity in “Homogeneous” Warehouse-Scale Computers: A Performance Opportunity

Abstract

The class of modern datacenters recently coined as "warehouse scale computers" (WSCs) has traditionally been embraced as homogeneous computing platforms. However, due to frequent machine replacements and upgrades, modern WSCs are in fact composed of diverse commodity microarchitectures and machine configurations. Yet, current WSCs are designed with an assumption of homogeneity, leaving a potentially significant performance opportunity unexplored. In this paper, we investigate the key factors impacting the available heterogeneity in modern WSCs, and the benefit of exploiting this heterogeneity to maximize overall performance. We also introduce a new metric, opportunity factor, which can be used to quantify an application's sensitivity to the heterogeneity in a given WSC. For applications that are sensitive to heterogeneity, we observe a performance improvement of up to 70% when employing our approach. In a WSC composed of state-of-the-art machines, we can improve the overall performance of the entire datacenter by 16% over the status quo.

Optimizing Google's warehouse scale computers: The NUMA experience

“…Very recently, researchers start to investigate the tradeoffs between improving NUMA performance and reducing cache contention [3,22]. In addition, related to constructing, scheduling and optimizing datacenters and WSCs, several prior work presents the challenges and proposed solutions [2,6,[12][13][14]16,21,23,24,30].…”

Mitigating Performance Unpredictability in Heterogeneous Clouds

“…Heterogeneity can be divided into two categories: hardware and workload heterogeneity. [14] evaluated the impact of assuming a homogeneous data center while it is heterogeneous. It proposes a metric to express an application sensibility facing heterogeneity.…”

REPP-H: Runtime Estimation of Power and Performance on Heterogeneous Data Centers

“…Managing the power and performance requirements of the applications is challenging because these data centers, incidentally or intentionally, have to deal with server architecture heterogeneity [19], [22]. One critical challenge that data centers have to face is how to manage system power and performance given the different application behavior across multiple different architectures.…”

“…However, these approaches require multiple iterations before power-performance criteria are satisfied and can lead to massive violations in powerperformance budgets for data centers [19], [27].…”