[sheepdog] Some thoughts based on my preliminary I/O test

[Liu Yuan]

On 05/23/2013 02:10 PM, Hongyi Wang wrote:
> I understand this is not a failure. But the problem is big IO which
> causes "poll timeout" hugely degrades the performance. For my test, I
> tried to write a 10GB file (bs=4m), which is expected to complete in 30m
> even in 100M network bandwidth. But I cannot finish it in 2 hours
> (actually I had to kill the task). That is a bit wield.

Umm, your emailer disrupt the threading (it was in a new thread instead
of the one it replied to), would be nice if you can fix this issue for a
better mail threading review.

how did you get 'expected to complete in 30m'? By the way bs=4m doesn't
guarantee underlying iosched will always issue the requests with 4M
requests to sheepdog, it is more of a hint.

Do you use object cache? Looks to me you don't use object cache.

Poll is means kernel is just waiting for the response without wasting
any cycles.

I am not sure why it takes longer than expected because it already take
full speed of network, it means requests are sending in a full speed.