I got a request from a reader to repeat the sysbench fileio benchmarks against the new Amazon EBS (Elastic Block Store).
As mentioned elsewhere as this is similar to network attached storage (NAS) it will be constrained by the speed of network.
The past series on sysbench fileio on EC2
Sysbench fileio vs EC2 Part 1
Sysbench fileio vs Large EC2 Part 2
seeker io benchmark vs EC2
As expected the network (1 Gigabit) is the upper limit of throughput (MB/s).
What this means is the local disk (not /mnt) is still the fastest disk available in terms of requests per sec and throughput (MB/s).
Adding an addition EBS volume as a RAID 0 was marginally better for single threaded reads but for many threads both the read performance was better.
The write performance was much better for single and many threads.
The additional flexibility of being able to attach and detach volumes, take snapshots is the big plus for using EBS.
At this point I made no attempt to optimize the RAID 0 chunksize. It would require more work, however at the end of the day, 90-100 MB/s is maximum.
How to use:
Same as before, stick your intensive IO portions on / and anything readonly on either /mnt or on Amazon EBS.
I am hanging out for an instance which provides infiniband network speeds or better. A bunch of clustering technology (shared nothing and shared disk) rely on good network bandwidth.
p.s. I had written this up about a month and half ago however delayed publishing as it isn't a great result i.e. network disk is constrained by network bandwidth.
Monday, September 1, 2008