Evaluation of a Virtual Xen Cluster Using the Pallas MPI Benchmarks Suite

Xuehai Zhang (hai@cs.uchicago.edu)
Katarzyna Keahey (keahey@mcs.anl.gov)

Last updated: 4/28/05

Pallas MPI Bechmarks or 'PMB' (now called Intel® MPI Benchmarks) is a suite of MPI benchmarks to measure the most important MPI functions. Please read the user manual to find the details of the benchmarks.

We use 8 PMB benchmarks to study the MPI performance on clusters under 4 case scenarios (We use the PMB-MPI1 part of PMB 2.2 packages for our study).

The 8 PMB benchmarks are PingPing, PingPong, SendRecv, Exchange, Bcast, Allgather, Alltoall, and Reduce. They can be classified to two catgories based on the communication pattern. PingPing, PingPong, SendRecv, and Exchange belong to Point-to-Point category while the rest benchmarks belong to the category of Collective Operations (as function of no. of CPUs).

The 4 case scenarios are:

1. linux: non-SMP native Linux cluster without Xen running. Each node has a single 500MHz CPU and 512MB RAM and runs kernel 2.4.29.
2. dom0: Xen dom0 cluster. Each Xen dom0 uses a 500MHz CPU and 128MB RAM and runs ported kernel 2.4.28.
3. domU(SMP): Xen domU virtual cluster. Each Xen domU runs on a different 500MHz CPU as dom0 and has 360MB RAM and a ported 2.6.10 kernel.
4. domU(noSMP): Xen domU virtual cluster. Each Xen domU runs on the same 500MHz CPU as dom0 and has 360MB RAM and a ported 2.6.10 kernel.

The following pages contain the performance results for all 8 benchmarks. We normalize the results of dom0,domU(SMP),and domU(noSMP) to those of linux for better illustration.

1. PingPing
2. PingPong
3. Sendrecv
4. Exchange
5. Bcast
6. Allgather
7. Alltoall
8. Reduce

Note: the value for each data point in the performance results (throughput or latency) is the average of 10 runs of the same benchmark.