0

I have tried benchmarking GlusterFS vs NFS on Amazon Web Services (AWS) using a m1.medium sized EC2 instance. We use the AWS EBS as the block store, with a XFS file system. Running Ubuntu 12.04 using the standard packages.

I used iozone and dd to do the benchmark. I realise these are not comparable as benchmarking tools, but I am getting some strange results.

My results in MB/s are below (all tests are run on the client):

using: iozone -c -e -i 0 -+n -r 64k -s 1000M -t 2

Direct to EBS     GlusterFS     NFS       GlusterFS + NFS
37.5              26.8          99.8      21.1

.

using: dd if=/srv/test of=/dev/null bs=64k count=16k

Direct to EBS     GlusterFS     NFS       GlusterFS + NFS
97.0              40.8          58.3      23.8

The direct to EBS doing well using dd, but with iozone it is actually slower than NFS. Why?

In general I have seen that GlusterFS is typically 1.5 to 2 times slower than NFS. Are these figures realistic for GlusterFS?

psiphi75
  • 151
  • 7

1 Answers1

1

Like all benchmarks it's difficult to compare apples with apples. It appears that iozone thrashes the system and doesn't simulate a specific load. dd just measures the throughput.

I came across filebench. With filebench you can simulate specific types of load, it calls these different tests "personalities". This gave me much more realistic numbers. I simulated a websever which typically is read only. Although filebench adds a log append to the simulation, I removed this in this case. I got the following results:

Results in MB/s

using: sudo filebench (webserver personality without simulation of log append)
Direct to EBS     GlusterFS     NFS       GlusterFS + NFS
24.2              6.0           17.0      20.4

Note: filebench was the constraint in the EBS test, since it was hogging 100% of the CPU during the test. I presume the EBS result would be higher.

Note 2: GlusterFS by itself had a very high latency when I included the test for log append. This may be explained in this post. I did not investigate this high latency issue since we expect mainly reads.

psiphi75
  • 151
  • 7