[rds-devel] comments on the send CQ completion handler

Or Gerlitz ogerlitz at voltaire.com
Thu Jan 10 07:06:06 PST 2008


Richard Frank wrote:
> If we are going to change this - let's run a test with out setting up 
> completion notifications for the rdma ops - and measure interrupt / 
> context switch rates / CPU util with and with out them.

Rick,

Please note that I don't suggest here to remove the IB_SEND_SIGNALED bit 
which is the one that causes a completion to be generated for the rdma 
work request but rather the IB_SEND_SOLICITED bit which causes a 
completion event to be generated in the receiver side, and since no 
completion is generated for rdma at the remote side, its useless and 
confusing to have it the code, that's all...

not asking for completions for the rdma work requests is something else.

Or.

> 
> I believe I did this in some past life - and it shows significant 
> reductions if we do not generate the extra completions. Who knows - 
> perhaps they are not needed at all - or maybe things have changed and 
> they are now !
> 
> Olaf Kirch wrote:
>> On Thursday 10 January 2008 12:33, Or Gerlitz wrote:
>>  
>>> - IB_SEND_SOLICITED is useless for rdma work requests as there is no 
>>> matching receive in the other side
>>>
>>> - max_unsig_bytes accounting need not be done for the rdma bytes 
>>> since its pages that came from the user and not allocated by rds
>>>
>>> Both are cleanups that improve the code's clarity, If you agree to 
>>> either/both - I will send a patch early next week
>>>     
>>
>> Yes, please do that. Cleanups always welcome :-)
>>
>> Thanks,
>> Olaf
>>   





More information about the rds-devel mailing list