[Openib-windows] [PATCH] WPP for al

[Yossi Leybovich]

> -----Original Message-----
> From: ftillier.sst at gmail.com [mailto:ftillier.sst at gmail.com] 
> On Behalf Of Fabian Tillier
> Sent: Friday, June 09, 2006 8:42 AM
> To: Yossi Leybovich
> Cc: openib-windows at openib.org
> Subject: Re: [Openib-windows] [PATCH] WPP for al
> 
> Hi Yossi,
> 
> I just ran some tests with the pingpong app over WSD, with 
> and without WPP in UAL, and the results for small messages is 
> that WPP adds about a third of a microsecond to latencies.  
> Note that I ran these test with the MT23108 driver, not the 
> MTHCA driver, because I wanted to isolate the WPP effects on 
> just one DLL.  With WPP enabled in the MTHCA UVP, I expect 
> things to be a little worse still.
> 
> We should plan to scrub the WPP calls and eliminate 
> superfluous entry/exit logging where possible.  I think 
> logging for the main entry point enter/exit is fine, as well 
> as errors.  I don't think we need to log entry/exit for every 
> sub-function of main API entrypoints.

I agree on that, for quick action we can remove all AL_ENTER/EXIT macros
from WPP.
I also think that the OBJ and HNDL bits add many prints we need to find
way to remove them as well.

> Depending on the feedback we get with the upcoming build, we 
> may have to disable WPP tracing in user-mode until we come up 
> with a less intrusive solution.
> 
This can be easily done in the user SOURCE file.
Just undef ENABLE_EVENT_TRACING=1 in FREEBUILD and you back to the old
fre version without any print.

> Also, would it be possible to define macros specifically for 
> errors that don't use the flags, and just the level?  This 
> would let us eliminate the ERR flag bit definition.
> 
Yes.
I will work on patch .

> Anyhow, attached are the two results for pingpong, with and 
> without WPP in UAL.
> 


> - Fab
>