mirror of
git://git.code.sf.net/p/sbcl/sbcl
synced 2026-09-10 07:26:40 -04:00
88b98e2d20
* Remove the create_thread_lock and thread setup semaphore. The start function is handed off in a GC-safe way and the creating thread is allowed to continue as soon as pthread_create() returns. * Retain the memory from up to 2 exited threads for reuse by new threads. This is not shared with the FOREIGN-THREAD memory pool, but perhaps could be. While this adds more conditionalization to both C and Lisp, it should be adaptable to OS_THREAD_STACK so that ultimately we can have only one way of performing post-mortem freeing instead of three. Benchmarks show about 6x to 7x faster lisp thread creation. Incidentally this fixes lp#1595699.
144 lines
6.5 KiB
Plaintext
144 lines
6.5 KiB
Plaintext
(defglobal *counter* 0)
|
|
(declaim (fixnum *counter*))
|
|
|
|
;; how many threads to make, not all at once, of course
|
|
(defparameter howmany-threads 20000)
|
|
|
|
(defun thread-lambda ()
|
|
(atomic-incf *counter*))
|
|
|
|
(defun test1 (&optional (n howmany-threads))
|
|
(setq *counter* 0)
|
|
(dotimes (i n)
|
|
(sb-thread:join-thread (sb-thread:make-thread #'thread-lambda)))
|
|
(assert (= *counter* n)))
|
|
|
|
(defun test2 (&optional (n howmany-threads))
|
|
(setq *counter* 0)
|
|
(let (prev-thread)
|
|
(dotimes (i n)
|
|
;; while starting the new thread, join the previously made.
|
|
(let ((new-thread (sb-thread:make-thread #'thread-lambda)))
|
|
(when prev-thread
|
|
(sb-thread:join-thread prev-thread))
|
|
(setq prev-thread new-thread)))
|
|
(sb-thread:join-thread prev-thread))
|
|
(assert (= *counter* n)))
|
|
|
|
(defun test3 (&optional (n howmany-threads))
|
|
(setq *counter* 0)
|
|
;; try making 10 at a time and waiting for those 10 before making more
|
|
(dotimes (i (/ n 10))
|
|
(let ((threads))
|
|
(dotimes (i 10)
|
|
(push (sb-thread:make-thread #'thread-lambda) threads))
|
|
(dolist (thread threads)
|
|
(sb-thread:join-thread thread))))
|
|
(assert (= *counter* n)))
|
|
|
|
Without :pauseless-threadstart
|
|
-------------------------------
|
|
|
|
$ perf stat ... --eval '(test1)' --quit
|
|
6,435.72 msec task-clock:u # 1.103 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
219,735 page-faults:u # 0.034 M/sec
|
|
1,484,016,696 cycles:u # 0.231 GHz
|
|
17,008,247,359 stalled-cycles-frontend:u # 1146.10% frontend cycles idle
|
|
525,232,002 instructions:u # 0.35 insn per cycle
|
|
# 32.38 stalled cycles per insn
|
|
145,098,224 branches:u # 22.546 M/sec
|
|
4,045,481 branch-misses:u # 2.79% of all branches
|
|
|
|
5.834005191 seconds time elapsed
|
|
0.684522000 seconds user
|
|
6.391622000 seconds sys
|
|
|
|
$ perf stat ... --eval '(test2)' --quit
|
|
6,383.26 msec task-clock:u # 1.182 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
220,625 page-faults:u # 0.035 M/sec
|
|
1,435,852,231 cycles:u # 0.225 GHz
|
|
16,839,228,290 stalled-cycles-frontend:u # 1172.77% frontend cycles idle
|
|
524,776,029 instructions:u # 0.37 insn per cycle
|
|
# 32.09 stalled cycles per insn
|
|
144,951,979 branches:u # 22.708 M/sec
|
|
4,018,908 branch-misses:u # 2.77% of all branches
|
|
|
|
5.401308431 seconds time elapsed
|
|
0.773771000 seconds user
|
|
6.235207000 seconds sys
|
|
|
|
$ perf stat ... --eval '(test3)' --quit
|
|
7,813.04 msec task-clock:u # 1.133 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
131,699 page-faults:u # 0.017 M/sec
|
|
1,477,424,012 cycles:u # 0.189 GHz
|
|
20,769,787,623 stalled-cycles-frontend:u # 1405.81% frontend cycles idle
|
|
525,748,318 instructions:u # 0.36 insn per cycle
|
|
# 39.51 stalled cycles per insn
|
|
145,134,345 branches:u # 18.576 M/sec
|
|
4,412,369 branch-misses:u # 3.04% of all branches
|
|
|
|
6.894458409 seconds time elapsed
|
|
0.641120000 seconds user
|
|
7.864134000 seconds sys
|
|
|
|
Conclusion: approximately 300 microseconds to start a thread
|
|
|
|
With :pauseless-threadstart
|
|
----------------------------
|
|
|
|
$ perf stat ... --eval '(test1)' --quit
|
|
1,541.91 msec task-clock:u # 1.068 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
633 page-faults:u # 0.411 K/sec
|
|
801,317,853 cycles:u # 0.520 GHz
|
|
3,751,130,586 stalled-cycles-frontend:u # 468.12% frontend cycles idle
|
|
352,170,311 instructions:u # 0.44 insn per cycle
|
|
# 10.65 stalled cycles per insn
|
|
81,884,155 branches:u # 53.106 M/sec
|
|
1,334,916 branch-misses:u # 1.63% of all branches
|
|
|
|
1.443664147 seconds time elapsed
|
|
0.425407000 seconds user
|
|
1.471034000 seconds sys
|
|
|
|
$ perf stat ... --eval '(test2)' --quit
|
|
1,249.01 msec task-clock:u # 1.550 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
573 page-faults:u # 0.459 K/sec
|
|
668,480,325 cycles:u # 0.535 GHz
|
|
3,086,418,550 stalled-cycles-frontend:u # 461.71% frontend cycles idle
|
|
376,966,208 instructions:u # 0.56 insn per cycle
|
|
# 8.19 stalled cycles per insn
|
|
95,790,671 branches:u # 76.694 M/sec
|
|
918,120 branch-misses:u # 0.96% of all branches
|
|
|
|
0.805563424 seconds time elapsed
|
|
0.325539000 seconds user
|
|
1.162641000 seconds sys
|
|
|
|
$ perf stat ... --eval '(test3)' --quit
|
|
1,308.09 msec task-clock:u # 1.455 CPUs utilized
|
|
0 context-switches:u # 0.000 K/sec
|
|
0 cpu-migrations:u # 0.000 K/sec
|
|
4,838 page-faults:u # 0.004 M/sec
|
|
660,514,151 cycles:u # 0.505 GHz
|
|
3,230,465,234 stalled-cycles-frontend:u # 489.08% frontend cycles idle
|
|
364,307,811 instructions:u # 0.55 insn per cycle
|
|
# 8.87 stalled cycles per insn
|
|
92,966,084 branches:u # 71.070 M/sec
|
|
730,714 branch-misses:u # 0.79% of all branches
|
|
|
|
0.899161220 seconds time elapsed
|
|
0.339090000 seconds user
|
|
1.227741000 seconds sys
|
|
|
|
Conclusion: 40 to 100 microseconds to start a thread
|