Skip to contentSkip to conversation

quell interactive feeding frenzy

RFC PATCH Al Boldi 3 participants 4 messages Apr 9, 2006 – Apr 11, 2006

lore
Share & tools
lore

Al Boldi

author
#0 lore raw
Quoted 10 lines · bert hubert
Quoted 3 lines · Mike Galbraith

Ok, unusable may be overstated. Nonetheless, that bit of code causes serious problems. It makes my little PIII/500 test box trying to fill one 100Mbit local network unusable. That is not overstated.

If you try to make a PIII/500 fill 100mbit of TCP/IP using lots of different processes, that IS a corner load.

I'm sure you can fix this (rare) workload but are you very sure you are not killing off performance for other situations?

This really has nothing to do w/ workload but rather w/ multi-user processing /tasking /threading. And the mere fact that the 2.6 kernel prefers kernel-threads should imply an overall performance increase (think pdflush).

The reason why not many have noticed this scheduler problem(s) is because most hackers nowadays work w/ the latest fastest hw available which does not allow them to see these problems (think Windows, where most problems are resolved by buying the latest hw).

Real Hackers never miss out on making their work run on the smallest common denominator (think i386dx :).

Thanks!

--
Al

#1 lore raw
Quoted 16 lines · Al Boldi
Quoted 10 lines · bert hubert
Quoted 3 lines · Mike Galbraith

Ok, unusable may be overstated. Nonetheless, that bit of code causes serious problems. It makes my little PIII/500 test box trying to fill one 100Mbit local network unusable. That is not overstated.

If you try to make a PIII/500 fill 100mbit of TCP/IP using lots of different processes, that IS a corner load.

I'm sure you can fix this (rare) workload but are you very sure you are not killing off performance for other situations?

This really has nothing to do w/ workload but rather w/ multi-user processing /tasking /threading. And the mere fact that the 2.6 kernel prefers kernel-threads should imply an overall performance increase (think pdflush).

Quoted 7 lines

The reason why not many have noticed this scheduler problem(s) is because most hackers nowadays work w/ the latest fastest hw available which does not allow them to see these problems (think Windows, where most problems are resolved by buying the latest hw).

Real Hackers never miss out on making their work run on the smallest common denominator (think i386dx :).

Please don't trim the cc list. I almost didn't see this, and I really do want to hear each and every opinion.

-Mike

author
#2 lore raw
Quoted 2 lines · bert hubert

In general, Linux systems are not maxed out as they will disappoint that way (like any system running with id=0).

top - 16:59:23 up 29 min, 0 users, load average: 993.49, 796.33, 496.21

Tasks: 1039 total, 1000 running,  39 sleeping,   0 stopped,   0 zombie
Cpu(s):  47.6% user,  52.4% system,   0.0% nice,   0.0% idle,   0.0% IO-wait
Mem:    125796k total,   123344k used,     2452k free,       64k buffers
Swap:  1020088k total,     9176k used,  1010912k free,     1752k cached
  PID  PR  NI  VIRT  RES  SHR SWAP S %CPU    TIME+  Command
 3946  28   0  2404 1460  720  944 R  5.8   0:14.78 top
 4219  37   0  1580  488  416 1092 R  3.0   0:00.45 ping
 4214  37   0  1584  480  408 1104 R  2.8   0:00.46 ping
 4196  37   0  1580  480  408 1100 R  2.5   0:00.45 ping
 4175  37   0  1584  488  416 1096 R  2.3   0:00.30 ping
 3950  37   1  1580  492  416 1088 R  2.0   0:08.77 ping
 4136  37   0  1580  488  416 1092 R  2.0   0:00.36 ping
 4158  37   0  1584  484  408 1100 R  2.0   0:00.35 ping
 4177  37   0  1580  480  408 1100 R  2.0   0:00.27 ping
 4180  37   0  1580  484  408 1096 R  2.0   0:00.33 ping
 4194  37   0  1580  480  408 1100 R  2.0   0:00.40 ping
 4199  37   0  1584  484  408 1100 R  2.0   0:00.47 ping
 4189  37   0  1580  488  416 1092 R  1.8   0:00.43 ping
 4153  37   0  1584  480  408 1104 R  1.5   0:00.31 ping
 4170  37   0  1584  484  408 1100 R  1.5   0:00.30 ping
 4191  37   0  1584  492  416 1092 R  1.5   0:00.41 ping
 4209  37   0  1584  484  408 1100 R  1.5   0:00.39 ping
 4215  37   0  1584  484  408 1100 R  1.5   0:00.39 ping
 4221  37   0  1580  492  416 1088 R  1.5   0:00.37 ping
 4146  37   0  1580  488  416 1092 R  1.3   0:00.29 ping
 4156  37   0  1584  484  408 1100 R  1.3   0:00.32 ping
 4166  37   0  1584  488  416 1096 R  1.3   0:00.33 ping
 4183  37   0  1580  480  408 1100 R  1.3   0:00.37 ping
 4216  37   0  1584  480  408 1104 R  1.3   0:00.39 ping
 4229  37   0  1584  484  408 1100 R  1.3   0:00.41 ping
 4233  37   0  1584  488  416 1096 R  1.3   0:00.41 ping
 4137  37   0  1580  484  408 1096 R  1.0   0:00.33 ping
 4141  37   0  1584  484  408 1100 R  1.0   0:00.31 ping
 4150  37   0  1580  484  408 1096 R  1.0   0:00.30 ping
 4161  37   0  1580  480  408 1100 R  1.0   0:00.29 ping
 4172  37   0  1584  484  408 1100 R  1.0   0:00.28 ping
 4178  37   0  1580  480  408 1100 R  1.0   0:00.25 ping
 4182  37   0  1584  484  408 1100 R  1.0   0:00.23 ping

After that the loadavg starts to wrap.
And even then it is possible to login.
And that's not with the default 2.6 scheduler, but rather w/ spa.

Thanks!

--
Al

#3 lore raw

Hi Al

Quoted 3 lines · Al Boldi

After that the loadavg starts to wrap.
And even then it is possible to login.
And that's not with the default 2.6 scheduler, but rather w/ spa.

Since you seem to use plugsched, I wonder if you could tell me how does current staircase perform with a load like that?

Signature · 1 line
-ck