Re: R: R: Ricompilazione kernel, performance
Alle 15:45, lunedì 11 giugno 2007, hai scritto:
> Altri vantaggi, li hai avuti?
> Io, per esempio, su un veeechio P2, con un kernel ottimizzato, ho
> dimezzato (cronometro alla mano) i tempi di boot.
Sicuramente è moooolto più veloce nel boot, ma ad occhio direi anche
nell'utilizzo normale. Ora sto giocando con un po di suite di benchmark tra
cui ltp e lmbench e sto avendo risultati impressionanti addirittura mi
sembrano esageratamente positivi. Tanto per farti un esempio:
Statistiche col kernel debian ufficiale
[lmbench3.0 results for Linux debian-etch 2.6.18-4-686 #1 SMP Wed May 9
23:03:12 UTC 2007 i686 GNU/Linux]
[...]
Simple syscall: 0.3245 microseconds
Simple read: 0.7405 microseconds
Simple write: 0.6664 microseconds
Simple stat: 3.4559 microseconds
Simple fstat: 1.0860 microseconds
Simple open/close: 5.0596 microseconds
Select on 10 fd's: 1.7285 microseconds
Select on 100 fd's: 8.4266 microseconds
Select on 250 fd's: 19.1701 microseconds
Select on 500 fd's: 37.9795 microseconds
Select on 10 tcp fd's: 1.8707 microseconds
Select on 100 tcp fd's: 19.5374 microseconds
Select on 250 tcp fd's: 53.3462 microseconds
Select on 500 tcp fd's: 98.1754 microseconds
Signal handler installation: 1.5952 microseconds
Signal handler overhead: 5.4904 microseconds
Protection fault: 0.7877 microseconds
Pipe latency: 9.8301 microseconds
AF_UNIX sock stream latency: 16.6156 microseconds
Process fork+exit: 240.1304 microseconds
Process fork+execve: 885.8333 microseconds
Process fork+/bin/sh -c: 3571.0000 microseconds
File /var/tmp/XXX write bandwidth: 15128 KB/sec
Pagefaults on /var/tmp/XXX: 2.2780 microseconds
[...]
Socket bandwidth using localhost
0.000001 0.56 MB/sec
0.000064 16.97 MB/sec
0.000128 32.50 MB/sec
0.000256 12.15 MB/sec
0.000512 101.94 MB/sec
0.001024 170.17 MB/sec
0.001437 167.28 MB/sec
10.485760 404.87 MB/sec
[...]
Memory write bandwidth
0.000512 2319.98
0.001024 2280.74
0.002048 2358.90
0.004096 2355.50
0.008192 2370.98
0.016384 2370.67
0.032768 2346.07
0.065536 1718.44
0.131072 1741.66
0.262144 1737.92
0.524288 1736.05
1.05 1673.85
2.10 1075.46
4.19 563.45
8.39 479.71
16.78 469.02
33.55 471.05
67.11 466.17
134.22 459.99
268.44 455.75
Mentre questi sono i valori col kernel ricompilato da me:
[lmbench3.0 results for Linux debian-etch 2.6.18-custom-asus #4 PREEMPT Mon
Jun 11 13:17:01 CEST 2007 i686 GNU/Linux]
[...]
Simple syscall: 0.1169 microseconds
Simple read: 0.2787 microseconds
Simple write: 0.2250 microseconds
Simple stat: 1.1811 microseconds
Simple fstat: 0.4082 microseconds
Simple open/close: 1.8092 microseconds
Select on 10 fd's: 0.6446 microseconds
Select on 100 fd's: 3.2200 microseconds
Select on 250 fd's: 11.1725 microseconds
Select on 500 fd's: 21.6314 microseconds
Select on 10 tcp fd's: 0.7141 microseconds
Select on 100 tcp fd's: 6.8853 microseconds
Select on 250 tcp fd's: 19.3732 microseconds
Select on 500 tcp fd's: 36.3553 microseconds
Signal handler installation: 0.5977 microseconds
Signal handler overhead: 1.6283 microseconds
Protection fault: 0.2758 microseconds
Pipe latency: 3.2447 microseconds
AF_UNIX sock stream latency: 5.2287 microseconds
Process fork+exit: 81.2537 microseconds
Process fork+execve: 334.3333 microseconds
Process fork+/bin/sh -c: 1283.4000 microseconds
File /var/tmp/XXX write bandwidth: 13950 KB/sec
Pagefaults on /var/tmp/XXX: 1.0599 microseconds
[...]
Socket bandwidth using localhost
0.000001 1.83 MB/sec
0.000064 52.00 MB/sec
0.000128 97.75 MB/sec
0.000256 37.42 MB/sec
0.000512 309.03 MB/sec
0.001024 502.35 MB/sec
0.001437 536.56 MB/sec
10.485760 792.39 MB/sec
[...]
Memory write bandwidth
0.000512 6168.50
0.001024 6202.56
0.002048 6265.50
0.004096 6303.02
0.008192 6288.78
0.016384 6302.22
0.032768 6274.22
0.065536 4578.36
0.131072 4625.79
0.262144 4619.69
0.524288 4608.92
1.05 4622.67
2.10 1439.61
4.19 623.13
8.39 517.50
16.78 513.74
33.55 504.49
67.11 504.06
134.22 491.13
268.44 469.73
536.87 458.08
Come puoi vedere in media ottengo un fattore di speedup pari a 3x il che
ripeto mi sembra addirittura esagerato...
Meglio esageratamente alto che il contrario, di certo dopo questi benchmark
non utilizzero mai più un kernel official, ma sempre e solo ricompilato ad
hoc :-) :-)
Saluti
--
Fabio Napoleoni
f.napoleoni@email.it
****************************************************************
"Computer Science is no more about computers than astronomy is
about telescopes"
Edsger W. Dijkstra
****************************************************************
Reply to: