[sldev] SL on Xeon processors

[Peter Phillips]

Dzonatas,

Are you alluding to the use of hyper-threading?  For what it's worth I 
wound up turning hyper-threading off on the dual xeon box I was using to 
build and run Second Life and found that it generally "felt" more 
consistent than having HT turned on.  I think my compile times slowed a 
bit, but overall the trade-off felt very worthwhile.

Cheers,
Peter

Dzonatas wrote:
> The Xeon processor appears to have something a bit different about it, 
> which the compile for mainline pentium 4 with GCC doesn't include. The 
> Windows OS appears to be a bit more random in how it schedules 
> processes on different parts of Xeon CPUs. It is obvious the software 
> will have to do a little more work to fix which threads get scheduled 
> on each logical unit. It is a bit easier to do on Windows. On Linux, a 
> special program would have to be run in order to properly enable 
> optimal schedules.
>
> The problem of the Xeon environment in default Windows can be easily 
> exploited with random benchmarks of different threads under heavy 
> cache hits. As SL starts up, the collection of threads that get 
> schedule on any particular logical unit is fairly arbitrary. A simple 
> solution would be to pair up those threads that can multi-thread 
> optimally together on a single logical unit.
>
> Don't think that SL is not capable of being able to run tests to 
> determine the best thread chain, but it can tend to look not 
> worthwhile to complete the analysis phase.
>
> Beyond SL, next years machine that have dynamic multi-core technology 
> have already started to claim such analysis phase worthwhile for many 
> application. Simply and recently, processor producers continue to 
> claim that most application still do not know how to handle multi-core 
> technology optimally.
>
> Just a helpful note. =)
>