[racket] Thread creation slow?

From: Matthew Flatt (mflatt at cs.utah.edu)
Date: Fri Aug 10 15:05:28 EDT 2012

I've pushed some improvements in v5.3.0.18.

On my machine (MacBook Air, 32-bit mode):

 v5.3:       27.274u 5.868s 0:35.14 94.2%   0+0k 6+2io 622pf+0w
 v5.3.0.18:   5.851u 2.494s 0:09.12 91.4%   0+0k 2+0io 691pf+0w

I believe that most of the savings are in corrections to avoid O(N)
setup for N threads for each minor GC.

Racket threads can be made lighter by adjusting the
`current-thread-initial-stack-size' parameter. If I change the program
to use

  (parameterize ([current-thread-initial-stack-size 100])
    (setup-whispers leftmost 0 100000))

then I get

              3.854u 1.175s 0:05.32 94.3%   0+0k 1+1io 689pf+0w

The trade-off for a smaller initial stack size is that deeper recursion
is more expensive.

Here are results in 64-bit mode on my machine:

 v5.3:         37.339u 24.869s 1:29.12 69.7%   0+0k 1+0io 49597pf+0w
 v5.3.0.18:     8.948u 10.030s 0:36.29 52.2%   0+0k 2+0io 43919pf+0w
 v5.3.0.18/100: 6.334u  3.317s 0:10.83 89.0%   0+0k 0+0io 0pf+0w

For the first two, there's a big difference between CPU+user time and
real time. Those configurations apparently use enough memory to
aggravate my machine's memory system.

Posted on the users mailing list.