The Who, What, Why and How of High Performance Computing Applications

The Who, What, Why and How of High Performance Computing Applications in the Cloud Soheila Abrishami 1

Introduction Cloud computing is emerging as an alternative to supercomputers for some of the high-performance computing (HPC) applications that do not require a fully dedicated machine Advantages of using cloud computing in HPC applications cost effective alternative reduces the risks caused by under-provisioning reduces the underutilization of resources caused by overprovisioning the built-in virtualization support in the cloud support flexibility, customization, security, migration and resource control 2

Motivation It still remains unclear whether, and when, HPC in the cloud can become a feasible substitute or complement to supercomputers. Based on the expansion of clouds to HPC applications, HPC users and cloud providers faced with the problem of choosing the best platform by having a limited knowledge about application characteristics, platform capabilities and cost why and who should choose (or not choose) cloud for HPC for what applications, and how should cloud be used for HPC 3

Contribution For answering what HPC applications are suitable for cloud By analyzing the performance of HPC applications on a range of platforms varying from supercomputer to cloud For answering how to use cloud for HPC By analyzing the impact of virtualization on HPC applications By using multiple platforms (dedicated and in the cloud) and use a smart application aware mapping of applications to platforms For answering why it is challenging to make a profitable business for clouds provider and also who can benefits from a HPC cloud By investigating the economic aspects of running in cloud vs. supercomputer 4

Experimental Testbed 5

Benchmarks and Applications Benchmarks and applications are selected from different scientific domains and which differ in nature, amount, and pattern of inter-processor communication. Benchmarks are chosen which are written in two different parallel programming environments - MPI and CHARM . These benchmark are: NAS Parallel Benchmarks (NPB) class B Jacobi2D NAMD ChaNGa Sweep3D NQueens 6

The scaling behavior of platforms for the selected applications 7

Performance Variation for ChaNGa 8

Latency and Bandwidth vs. Message Size on all platforms 9

System noise has detrimental impact on performance 10

Optimizing Cloud Virtualization for HPC To mitigate the overhead of cloud platform two ways is considered Lightweight virtualization : reduces the latency overhead of network virtualization by granting virtual machines native accesses to physical network interfaces CPU affinity: instructs the operating system to bind a process (or thread) to a specific CPU core 11

Lightweight virtualization Two lightweight virtualization techniques thin VMs configured with PCI pass-through for I/O containers that is OS-level virtualization Thin VM: Definition: a physical network interface is allocated exclusively to a thin VM Problem: under utilization when the thin VM generates insufficient network load Containers: Definition: share the physical network interface with its sibling containers and its host. Problem: containers must run the same operating system as their underlying host 12

Impact of Virtualization on Application Performance 13

CPU Affinity Prevent migration of process Prevent multi process share a core Improving cache locality In cloud CPU affinity can be enforced in two levels: Application level: binding process of to the virtual CPUs of a VM Hypervisor level: binding virtual CPUs to physical CPUs 14

Impact of CPU Affinity on CPU Performance operations 12 process 15

Application Performance with various CPU Affinity Settings, using thin VM and plain VM 16

HPC Economic in the Cloud Why it is challenging to make a profitable business for cloud providers for HPC: Utilization of resources on HPC system is high HPC cloud user would want a dedicate instance The performance of HPC applications is very sensitive to the interconnect 17

Cost ratio of running in cloud and a dedicated supercomputer by assuming different per-core-hour cost ratio from 1x to 5x 18

Cloud Bursting and Benefit How to find a mapping between application and available resources to makes best use of the dedicated HPC-optimized resources, minimizes the cost of bursting to the cloud, and meet the performance targets? By an intelligent mapping algorithms Which aware of application characteristics Understand that application which scale poorly on cloud should be allocated to dedicated resources first 19

parallel efficiency where is the number of processors, and Speedup is defined as: where is the sequential execution time and is the parallel execution time Mapping of HPC applications to platforms with varying resources (e.g., different processor types and speed, interconnection networks, and virtualization overhead) 20

Applications with high parallel efficiency are good candidates for cloud 21

Using of smart mapping algorithm in different scenarios A smart mapping algorithm by using the application signature and platform characteristic can recommend the best platform for a given application under different scenarios Minimize Cost with Performance Guarantees Maximize Performance with Constrained Budget 22

Cost with Performance Guarantees 23

Summary Lightweight virtualization is important to remove overheads for HPC in cloud. A hybrid cloud-supercomputer platform environment can outperform its individual constituents Application characterization in the HPC-cloud space is challenging but the benefits are substantial 24