Speaker
Description
Summary
During the past decade public clouds have attracted tremendous amount of interest from academic and industrial audiences as the effective and relatively cheap way to get powerful computational infrastructure without the burden of building and maintaining physical infrastructure.
Although clouds are less powerful than server clusters or supercomputers [1], they are becoming more popular as a platform for High Performance Computing (HPC) due to the low cost and easy to access. Cloud providers are starting to support this interest and come up with a new cloud paradigm - HPC-as-a-service. This paradigm represents a service that gives cloud resources for computationally heavy applications.
Several papers [2, 3] have shown that one of the main performance bottlenecks in HPC clouds issues from communication delays within the DС network. Such bottleneck is due to the insufficient network performance in HPC clouds. While supercomputers use fast interconnections like InfiniBand or GE, HPC clouds mostly use Ethernet.
However, this bottleneck brings important impact on the behavior of the applications in HPC clouds – communication heavy HPC applications tend to underutilize the CPU. This happens because most of computationally heavy applications use network to exchange messages between physical machines. And since cloud network is not fast enough for HPC, such applications spend a lot of time idling for messages to pass through the network [2, 3].
Such behavior of HPC applications also leads to a highly regular execution and network usage pattern, i.e. HPC applications show tendency to alternate computations with frequent network communications [5]. This communication pattern contribute to the idle CPU usage since the slowest message delivery dictates the overall performance degradation of an application.
These specifics of the behavior of HPC applications can be used in HPC clouds to improve the resource utilization by sharing the same CPU core between different applications, i.e. providing more virtual CPUs than there are physical ones. In this research we are proposing a scheduling algorithm that increases the resource utilization and the HPC task capacity of an Ethernet-based HPC cloud. The developed algorithm observes network behavior of HPC tasks and uses a greedy heuristic to share CPU cores between such tasks, thus improving the overall CPU usage and increasing the number of tasks performed via HPC-as-a-service.
We have performed experiments with 15 popular MPI benchmarks/libraries and show that we can significantly improve CPU usage with negligible performance degradation.
[1] Netto, M. A., Calheiros, R. N., Rodrigues, E. R., Cunha, R. L., & Buyya, R. (2018). HPC cloud for scientific and business applications: Taxonomy, vision, and research challenges. ACM Computing Surveys (CSUR), 51(1), 8.
[2] Gupta, A., Faraboschi, P., Gioachin, F., Kale, L. V., Kaufmann, R., Lee, B. S., ... & Suen, C. H. (2016). Evaluating and improving the performance and scheduling of HPC applications in cloud. IEEE Transactions on Cloud Computing, 4(3), 307-321.
[3] Gupta, A., & Milojicic, D. (2011, October). Evaluation of hpc applications on cloud. In 2011 Sixth Open Cirrus Summit (pp. 22-26). IEEE.
