[sheepdog-users] 50.000 iops per VM

Liu Yuan namei.unix at gmail.com
Wed Jul 4 08:28:47 CEST 2012

On 07/04/2012 02:15 PM, Christoph Hellwig wrote:
> On Wed, Jul 04, 2012 at 01:59:46PM +0800, Liu Yuan wrote:
>> I am writing some patches to offload the data transfer work to threads,
>> with this change, I think the performance will be boosted. But I think
>> those patches are queued after 0.4.0 release.
> Interesting, I haven't really seen non-blocking event based I/O causing
> a bottleneck that easily.

It is only my guess, maybe not that accurate but I wanna give it a try.
I think it is non-blocking related, but rather current implementation
bundles notification(poll) with data transfer all in main thread. Then
both are heavily affecting each other.

> What model do you plan to use?  Just a separate TX thread, or some
> sort of pool that will require additional locking?

I just separate notification and actual data transfer, so TX and RX are
offloaded to worker threads. I think thread pool or directly use short
thread is irrelevant, it is how we implement worker threads.


More information about the sheepdog-users mailing list