Skip to contentSkip to conversation

RFA: Changing scheduler quantum (Was: REQUEST: OpenLDAP 2.3.7)

DISCUSSION Bernardo Innocenti 2 participants 4 messages Sep 18, 2005 – Sep 19, 2005 6 messages not in the archive

lore
Share & tools
lore
#6 lore raw
In reply to a message not in the archive
Quoted 6 lines · Arjan van de Ven
Quoted 2 lines · Bernardo Innocenti

It's more meaningful to interpret sched_yield() as "give up the processor, as if the scheduler quantum had expired".

afaik this is *exactly* what the new sched_yield() does ;)

Oops :-)

Quoted 5 lines
Quoted 3 lines

The scheduler wouldn't normally allow a lower priority process to preempt a high-priority ready process for 30+ ms. Unless I'm mistaken about Linux's scheduling policy...

if your quantum is up... all other tasks get theirs of course

I assumed dynamic priorities affected the length of the quantum, but maybe it just changes the number of times the process is scheduled wrt other processes, with the quantum being fixed at 20-30ms.

(...a few seconds later...)

Skimming through sched.c, it seems my first guess was
right: the quantum varies with the priority from 5ms to 800ms.

The DEF_TIMESLICE of 400ms looks a bit too gross for most applications and the maximum 800ms is just ridicolously high.

IIRC, the 7.14MHz 68000 in the Amiga 500 did task-switching at 20ms intervals, with a negligible performance hit. Couldn't do much better on today's CPUs?

Signature · 2 lines
  // Bernardo Innocenti - Develer S.r.l., R&D dept.
\X/  http://www.develer.com/
#7 lore raw
Quoted 26 lines · Bernardo Innocenti
Quoted 5 lines · Arjan van de Ven
Quoted 2 lines · Bernardo Innocenti

It's more meaningful to interpret sched_yield() as "give up the processor, as if the scheduler quantum had expired".

afaik this is *exactly* what the new sched_yield() does ;)

Oops :-)

Quoted 5 lines
Quoted 3 lines

The scheduler wouldn't normally allow a lower priority process to preempt a high-priority ready process for 30+ ms. Unless I'm mistaken about Linux's scheduling policy...

if your quantum is up... all other tasks get theirs of course

I assumed dynamic priorities affected the length of the quantum, but maybe it just changes the number of times the process is scheduled wrt other processes, with the quantum being fixed at 20-30ms.

(...a few seconds later...)

Skimming through sched.c, it seems my first guess was
right: the quantum varies with the priority from 5ms to 800ms.

Quoted 7 lines

The DEF_TIMESLICE of 400ms looks a bit too gross for most applications and the maximum 800ms is just ridicolously high.

IIRC, the 7.14MHz 68000 in the Amiga 500 did task-switching at 20ms intervals, with a negligible performance hit. Couldn't do much better on today's CPUs?

Not quite.

The default timeslice of nice 0 tasks is 100ms. The timeslice is not altered the way you have read sched.c. It is altered thus:
1. For 'nice' levels it varies from 5ms at nice 19 to 800ms at nice -20.
2. For interactive tasks, it is cut up into smaller pieces down to 10ms and round robins with other tasks at the same dynamic priority, but still is based on the nice levels for the full length of cpu time before expiration overall.

Cheers,
Con

#8 lore raw
Quoted 8 lines · Con Kolivas
Quoted 3 lines · Bernardo Innocenti

The DEF_TIMESLICE of 400ms looks a bit too gross for most applications and the maximum 800ms is just ridicolously high.

Not quite.

Quoted 7 lines

The default timeslice of nice 0 tasks is 100ms. The timeslice is not altered the way you have read sched.c. It is altered thus:
1. For 'nice' levels it varies from 5ms at nice 19 to 800ms at nice -20.
2. For interactive tasks, it is cut up into smaller pieces down to 10ms and round robins with other tasks at the same dynamic priority, but still is based on the nice levels for the full length of cpu time before expiration overall.

I see. Then there must be something else to explain the behavior I'm observing with slapd.

Each and every call to sched_yield() makes the process sleep for over *50ms* while a "nice make bootstrap" is running in the background:

[pid  8780]      0.000033 stat64("gidNumber.dbb", 0xb7b3ebcc) = -1 EACCES (Permission denied)
[pid  8780]      0.000059 pread(20, "\0\0\0\0\1\0\0\0\1\0\0\0\0\0\0\0\0\0\0\0\2\0\344\17\2\3"..., 4096, 4096) = 4096
[pid  8780]      0.000083 pread(20, "\0\0\0\0\1\0\0\0\4\0\0\0\3\0\0\0\0\0\0\0\222\0<\7\1\5\370"..., 4096, 16384) = 4096
[pid  8780]      0.000078 time(NULL)    = 1124322520
[pid  8780]      0.000066 pread(11, "\0\0\0\0\1\0\0\0\250\0\0\0\231\0\0\0\235\0\0\0\16\0000"..., 4096, 688128) = 4096
[pid  8780]      0.000241 write(19, "0e\2\1\3d`\4$cn=bernie,ou=group,dc=d"..., 103) = 103
[pid  8780]      0.000137 sched_yield( <unfinished ...>
      ...zzzz...
[pid  8781]      0.050020 <... sched_yield resumed> ) = 0
[pid  8780]      0.000025 <... sched_yield resumed> ) = 0
[pid  8781]      0.000060 futex(0x925ab20, FUTEX_WAIT, 33, NULL <unfinished ...>
[pid  8780]      0.000026 write(19, "0\f\2\1\3e\7\n\1\0\4\0\4\0", 14) = 14
[pid  8774]      0.000774 <... select resumed> ) = 1 (in [19])

Actually, I'm now noticing that several slapd threads were involved here. Depending how strace handles relative timestamps of multiple processes, it may mean both 8780 and 8781 slept too much or just 8781 did and 8780 was quick.

Any idea? I'm planning to patch my kernel to print the time_slice value in /proc/*/stat. This way I can check it's being computed as intended for both slapd and gcc.

Signature · 2 lines
  // Bernardo Innocenti - Develer S.r.l., R&D dept.
\X/  http://www.develer.com/
#9 lore raw
Quoted 8 lines · Bernardo Innocenti
Quoted 7 lines · Con Kolivas
Quoted 3 lines · Bernardo Innocenti

The DEF_TIMESLICE of 400ms looks a bit too gross for most applications and the maximum 800ms is just ridicolously high.

Not quite.

Quoted 7 lines
Quoted 7 lines

The default timeslice of nice 0 tasks is 100ms. The timeslice is not altered the way you have read sched.c. It is altered thus:
1. For 'nice' levels it varies from 5ms at nice 19 to 800ms at nice -20.
2. For interactive tasks, it is cut up into smaller pieces down to 10ms and round robins with other tasks at the same dynamic priority, but still is based on the nice levels for the full length of cpu time before expiration overall.

Please do not cc mailing lists that reply with the "your email is awaiting moderator approval" to lkml.

Quoted 6 lines

I see. Then there must be something else to explain the behavior I'm observing with slapd.

Each and every call to sched_yield() makes the process sleep for over *50ms* while a "nice make bootstrap" is running in the background:

Why this preoccupation with how long sched_yield takes? We've already established that it takes a variable unpredictable (yet long) time for SCHED_NORMAL tasks. No, cancel that question or we'll start having people tell us what the kernel should do all over again.

You're almost certainly seeing the effect of fork during 'make bootstrap' and multiple tasks are running prior to expiration on the active runqueue. SCHED_NORMAL tasks that have done sched_yield will yield till nothing is left wanting cpu time on the active runqueue.

Quoted 8 lines

Actually, I'm now noticing that several slapd threads were involved here. Depending how strace handles relative timestamps of multiple processes, it may mean both 8780 and 8781 slept too much or just 8781 did and 8780 was quick.

Any idea? I'm planning to patch my kernel to print the time_slice value in /proc/*/stat. This way I can check it's being computed as intended for both slapd and gcc.

Feel free to do as much checking on kernel code as you like.

Cheers,
Con