Recommendations for Virtualization Technologies in High Performance Computing

Recommendations forVirtualization Technologies inHigh Performance Computing

Summarized by: Michael Riera

9/17/2011

University of Central Florida – CDA5532

Papers in discussion

• Recommendations for Virtualization Technologies in HighPerformance Computing– Nathan Regola, Center for Research Computing

• Notre Dame

– Jean-Christophe Ducom, Center for Research Computing,• Notre Dame

• The Architecture of Virtual Machines• The Architecture of Virtual Machines– James E. Smith

• University of Wisconsin-Madison

– Ravi Nair• IBM T.J. Watson Rsearch Center

• Understanding Performance Interference of I/O Workload invirtualized Cloud Environments

• Xing Pu, Ling Liu, Yiduo Mei, Sankaran Sivathanu, Younggyun Koh, Calton Pu– Georgia Tech, Beijing Institute of Technology, P.R. China, Xi’an Jiaotong University, P.R.

China

Agenda

• Purpose

• OpenVZ, KVM, Xen, EC2 (Running Xen)

• Benchmarks

• Experiment Setup• Experiment Setup

• Evaluation & Results

Purpose

• This paper evaluates three well known sourceVMM (hypervisors) on there disk throughput(read, write ops), round trip latency, andrelative running time.relative running time.

– OpenVZ

– KVM

– Xen (including Amazon EC2)

OpenVZ

• Is a modified Linux kernel that is enabled toact as a VMM

– Allows Linux based OS to run as processes

– One Kernel on the host to support both host and– One Kernel on the host to support both host andclient

• Single Point of failure for guest and host failures

– Lower CPU overhead because each VM is aprocess

KVM

• 10,000 Lines of Code that turns a Linuxdistribution into a VMM by loading a kernelmodule.

– Leveraging existing Linux infrastructure– Leveraging existing Linux infrastructure

• I/O Stack, Device Drivers, memory manager, scheduler

– Enables Guess Mode for other Linux kernels

– Virtio driver

• Enables para-vitualization- therefore allowing VM to requestaccess for computations to be made on a host computer.

• Enables PCI passthrough

Xen (including EC2)

• Loosely coupled from the VM running

– Supports Multiple OS distributions, including Solaris,Windows, BSD, and Linux

– VM Domain Isolation– VM Domain Isolation

• Cannot be used to attack an operating system; e.g. Xencannot attack the host operating system as there is no hostoperating system to attack (outside the memory space)

– Privileged Access

– Small Code Base

– Supports Para-virtualization

Benchmarks

• Auction Benchmarks through RUBIS– For High enterprise workload

• NAS Parallel benchmark (NPB)– A suite that emphasizes a particular type of numerical

computation and reproduces the CPU, cache, memory, and I/Osystem workload of a wide range of real world apps.system workload of a wide range of real world apps.

• Consist of five kernels (workloads):– EP, MG, CG, FT, IS

• Consist of three complex fluid dynamics (CFD) applications– BT, SP, LU

• OpenMP Benchmark– OpenMP Microbenchmarks

• MPI Benchmark– Intel MPI Benchmark 3.2.2

Experiment Setup

• Consist of:– Four identical Dell R610 machines, each with two 2.27 Ghz Intel

Xeon E5520 processor (quad-core)• 24GB of RAM• GB Nic card and Infiniband Qlogic 7240• Host Operating system: Fedora 12

VMM: OpenVz, KVM, Xen• VMM: OpenVz, KVM, Xen• Host VM on OpenVz

– 13 Linux Kernel 2.6.18

• Host VM on KVM– 13 Linux Kernel 2.6.32

• Host VM on Xen 3.4.3-2– 13 Linux kernel between 2.6.32-1.2.108.

• Guess virtual machines operating system– Redhat 5.4

Evaluation and Results

KVM Leads because of the para-virtualization (for reads/ random), butdoesn’t help for writes and random writes

Evaluation and Results

EC2 native Hyperthreading

Evaluation and Results

Evaluation and Results

Evaluation and Results