Poor I/O Performance

[dbgrandi]

I have been running some basic IO tests on a 1GB Ubuntu, single CPU, instance and noticed that the IO seems quite poor compared to some other machines I have tested against.

I used a combination of bonnie++ and iozone on 4 machines:
- my macbook laptop
- a flexiscale instance
- a dedicated server
- a VPS at slicehost.com

The flexiscale host is the slowest by far. I'm still analyzing the data of the three other machines and waiting for iozone to finish on our flexiscale instance...while the same test finished on my development laptop about 45 minutes ago (started at the same time.)

Are there any flexiscale representatives who can comment on the IO system and what we can do to get better performance. I realize that there is always a trade off between the convenience of being able to start/stop virtual servers on demand and performance, but I'm wondering what to expect as Flexiscale (as a company) grows and gets more customers in their datacenter.

-dave

[tonylucas]

Dave,

We are currently seeing some I/O issues which mean that performance for some customers is not where we would like it to be. We are hoping to see an improvement in this in the next 72 hours or so, as we do some maintenance work on the backend storage to improve this.

Apologies for the delayed reply on this.

Regards,

Tony.

[szferi]

I did some early test and I found that the I/O performance is not so impressive:
11 MB/s average linear write and 12.1 MB/s average linear read.

I used the dd with 1M block size for the tests and an instance is a Debian Etch 64bit instance with 20GB disk and 0.5 MB RAM.

Do you have any suggestion how I can increase I/O performance in current configuration?

Is there any plan to increase the I/O performance?

[tonylucas]

Jayson,

We are aware of specific issues both in transactions for servers (starting, stopping etc) and I/O performance.

Both issues are being worked on, and we do intend to deliver improvements in these areas as soon as possible.

We have long term plans to allow the system to scale far more easily to around 500 times it's current size, and are also taking steps in the short term to improve these.

In regards to the IO, the disk setup is actually far more complex than you have guessed, but some of the speed problems are due to contention right at the physical disk level, which we have an ongoing process to improve.

I apologise for any problems you are experiencing with it in this regard, but I'm also very pleased you are happy with everything else!

Please be assured we are very aware of what and where the problems are, and how to tackle them, so please do let us know how your testing goes.

Regards,

Tony Lucas
Chief Executive Officer
FlexiScale