[Gluster-devel] general performance translator recommendations

Brent A Nelson brent at phys.ufl.edu
Wed Jan 14 19:20:34 UTC 2009


On Wed, 14 Jan 2009, Raghavendra G wrote:

> Hi Brent,
>
> On Wed, Jan 14, 2009 at 6:56 PM, Brent A Nelson <brent at phys.ufl.edu> wrote:
>
>> On Wed, 14 Jan 2009, Vikas Gorur wrote:
>>
>>  As for io-threads above/below unify, there is not much to be gained by
>>> using io-threads on the client side, since there are no slow
>>> operations on the client side (just a straight code path from the
>>> application to the server). Using it on the server makes sense because
>>> it lets GlusterFS do other useful work while one thread is waiting for
>>> a read/write from/to the underlying filesystem finishes.
>>>
>>>
>> I found io-threads on the client to be a major gain.  When the client is
>> very busy with io tasks, io-threads seems to eliminate latency in metadata
>> operations (e.g., ls doesn't hang for ages).  It made the client feel much
>> more responsive.
>
>
> Yes. But with 2.0, Non-blocking I/O is introduced. Hence the client thread
> will not be blocked on I/O operations.
>

My testing was with 1.4 RCs; it was still of substantial benefit to 
metadata operation latency, when the client is busy with multiple 
concurrent streaming I/O tasks.

Thanks,

Brent





More information about the Gluster-devel mailing list