Hello J-beta team,
A brief setup below explores the interaction between multi-threaded
primitives, threads, and pools. There are two points and associated
questions.
As additional background for this, adding threads across pool 0 behaves as
expected: timing monotonically decreases with the addition of threads.
NB. 1) =========================================================
NB. Setup and a timing function for matrix multiplication primitive.
NB. Pool setup
syspools =. (8&T.)@(''&[)
delpools =. ([(55&T.))^:(1&T.)@(''&[) NB. pool clear: delpools y clears all
worker threads and pools
getpools =. ((2&T.)"0)@i. NB. pool stats: 'getpools 6' means get stats on
first six pools
setpools =. getpools@((#@;)"_)@:({{([(0&T.))^:({: y)
({.y)}}"1)@(,.~(i.@#))@([delpools)
NB. pool threadcount: 'setpools 0 2 2 3' creates 0 threads in pool 0, 2
thread in pool 1, ...
fmf =. {{1.01 * (y,y) $ ? (y^2) $ 2^62}}"0
'A B' =: fmf 2 # 1024
f =. {{(20) 6!:2 'A +/ . * B'}} NB. Matrix multiplication timing
uf =. {{u t. (y) ''}}"0 NB. Run u in threadpool y
NB. =========================================================
NB. Demonstrate multi-threading on primitives is oddly cyclical.
setpools 4
4 0 4
res =. > (((f uf)@0:))"0 (100 # 0)
syspools '' NB. 6 threads, 12 hyperthreads, in case this is related. Seems
like it shouldn't be.
12 63
(+/%#) _6 ]\ res NB. Only every 6th run appears to use multi-threaded
primitives
0.121716 0.121377 0.121108 0.121901 0.11124 0.0308138
*NB. Questions: What's the underlying behavior? Is this intended?*
NB. 2) =========================================================
NB. Docs say: "Parallelizable primitives use threadpool 0."
NB. Neither pool 0 or 1 appear to use multi-threaded primitives
setpools 0 5
0 0 0
5 0 5
> (f uf) 0,1 NB. Run timing for matrix multiplication in pools 0 and 1
0.118203 0.11434
NB. Both pools 0 and 1 appear to use multi-threaded primitives
setpools 4 1
4 0 4
1 0 1
> (f uf) 0,1 NB. Run timing for matrix multiplication in pools 0 and 1
0.0609 0.062134
NB. The docs seem to describe the threadpool model completely, but this
seems like an add model.
*NB. Questions:*
NB. Why wouldn't each pool use its own threads for multi-threaded
primitives?
NB. What's the point of creating a second pool with multiple threads,
rather than separate pools?
NB. How is resource contention between pools adjudicated?
Best,
John
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
