Re: [Aoetools-discuss] ggaoed and multiple NICs
Brought to you by:
ecashin,
elcapitansam
From: Gabor G. <go...@di...> - 2011-07-07 19:35:14
|
On Thu, Jul 07, 2011 at 06:24:00PM +0200, Alexandre Chapellon wrote: > I personnaly do not use the merge-delay as it is reported to be > experimental. Further more. All my tests demonstrated worth > performances with merge-delay (I also tried with 0.002 as you did). The problem with merge-delay is that its effect absolutely depends on your workload. As a rule of thumb, if most of the read/write requests generated on the client are large, then merge-delay may improve performance. On the other hand, increasing merge-delay will reduce the performance of small, random requests (either read or write). If you want to have a semi-scientific formula for calculating merge-delay, then you should measure how many network packets your average request generates, and calculate/measure how much time the server needs to receive those packets (from the time of reporting the first packet to the time of reporting the last). This depends on saturated your network is, how good your NIC is, does the driver use NAPI and if yes then what are the coalesce parameters etc. It may be a nice topic for a scientific paper, but in practice, experimentation is more effective. The above also means that even if you find a certain value good for you, it may be totally wrong for someone else with different server hardware, different network setup or different I/O load. I think this explains quite well why merge-delay is experimental :-) > That said my setup is not perfect and I don't want to misslead you. > I have tried several times to get in touch with Gabor Gombas who > wrote ggaoed... but unfortanately all my mails remained dead > letters. Sorry, I do not use AoE anymore, and I have little time for reading mailing lists, and even less for answering. Gabor |