yade-dev team mailing list archive

[Jan Stránský <_honzik@xxxxxxxxxx>]

Hello,

>We know that parallelized engines have usually no use of more than ~4-5 
>threads currently.
>Here, I have easily 10 proc. available for routine jobs. Which means I 
>could solve my linear system on, say, 4 procs while interactions are 
>dispatched on 4 other procs (also applies if only one engine is 
>parallelized).
Sorry, I live in relations of my 2 cores laptop :-) but still overall speedup for 8 cores is 5.5x (page 66 of Vaclav's thesis), so this, depending on engine, is not "of no use" yet, at least from my point of view.. but yes, parallel 4+4 at the same time would be better than 8 and then 8..

>Of course, I have to wait for the slowest one, but max(t1,t2) is always 
>better than t1+t2, no?
Of course :-) but max(2*t1/nCores,2*t2/nCores) > (t1/nCores+t2/nCores)  on 2 cores, on more cores probably not always

>With massive parallelism available is some algebra packages, we could 
>virtually run coupled simulation at almost the same speed as uncoupled 
>ones, provided the number of cores is high enough to solve the fluid 
>problem faster than interaction dispatch. :-)
In this case this is definitely the right approach :-)

Jan