Subject: Re: Port benchmarks
To: None <netbsd-ports@netbsd.org, port-alpha@netbsd.org>
From: Manuel Bouyer <bouyer@antioche.lip6.fr>
List: port-alpha
Date: 04/04/2000 13:29:43
On Mon, Apr 03, 2000 at 12:00:21PM +0200, Manuel Bouyer wrote:
>
> Ok, I've got new results, that I've merged with Simon's sumary.
> ds20-1 is with lmbench compiled '-O2 -mcpu=21264', ds20-2 is with lmbench and
> libc compiled '-O2 -mcpu=21264'.
> Note that a libc compiled this way is not fully functionnal, portmap gets
> a 'invalid instruction' in libc.
Ok, now that Jason fixed the compiler I've been able to compile a whole
system (kernel+userland) -O2 -mcpu=21264. Here are the results (ds20 NetBSD
1.4X), merged with the previous ones.
L M B E N C H 1 . 9 S U M M A R Y
------------------------------------
(Alpha software, do not distribute)
Processor, Processes - times in microseconds - smaller is better
----------------------------------------------------------------
Host OS Mhz null null open selct sig sig fork exec sh
call I/O stat clos inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ----- ---- ---- ---- ---- ----
ds20 NetBSD 1.4U 498 0.6 2.5 12 16 0.05K 1.2 4 0.7K 3K 5K
ds20-1 NetBSD 1.4U 498 0.6 2.8 13 17 0.05K 1.2 4 0.7K 3K 5K
ds20-2 NetBSD 1.4U 498 0.6 2.7 13 18 0.06K 1.3 4 0.7K 3K 5K
d20 NetBSD 1.4X 498 0.6 2.5 11 18 0.06K 1.3 4 0.7K 3K 5K
alpha-dec OSF1 T5.0 498 0.5 1.0 76 84 0.13K 0.9 3 0.8K 3K 6K
Context switching - times in microseconds - smaller is better
-------------------------------------------------------------
Host OS 2p/0K 2p/16K 2p/64K 8p/16K 8p/64K 16p/16K 16p/64K
ctxsw ctxsw ctxsw ctxsw ctxsw ctxsw ctxsw
--------- ------------- ----- ------ ------ ------ ------ ------- -------
ds20 NetBSD 1.4U 3 10 48 18 56 22 61
ds20-1 NetBSD 1.4U 4 12 48 20 61 25 66
ds20-2 NetBSD 1.4U 4 12 48 22 55 24 67
ds20 NetBSD 1.4X 5 12 49 25 67 26 74
alpha-dec OSF1 T5.0 3 8 41 16 42 17 45
*Local* Communication latencies in microseconds - smaller is better
-------------------------------------------------------------------
Host OS 2p/0K Pipe AF UDP RPC/ TCP RPC/ TCP
ctxsw UNIX UDP TCP conn
--------- ------------- ----- ----- ---- ----- ----- ----- ----- ----
ds20 NetBSD 1.4U 3 16 15 48 49 938
ds20-1 NetBSD 1.4U 4 16 24 49 50 961
ds20-2 NetBSD 1.4U 4 16 15
ds20 NetBSD 1.4X 5 20 19 50 49 1182
alpha-dec OSF1 T5.0 3 20 24 46 48
File & VM system latencies in microseconds - smaller is better
--------------------------------------------------------------
Host OS 0K File 10K File Mmap Prot Page
Create Delete Create Delete Latency Fault Fault
--------- ------------- ------ ------ ------ ------ ------- ----- -----
ds20 NetBSD 1.4U 1408 854 1515 2325 1785192 0.1K
ds20-1 NetBSD 1.4U 1612 833 2083 1639 1774413 0.1K
ds20-2 NetBSD 1.4U 1449 847 1818 1428 1775389 0.1K
ds20 NetBSD 1.4X 1515 833 1388 1754 430663 0.1K
alpha-dec OSF1 T5.0 63 120 276 182 1348 2 0.0K
*Local* Communication bandwidths in MB/s - bigger is better
-----------------------------------------------------------
Host OS Pipe AF TCP File Mmap Bcopy Bcopy Mem Mem
UNIX reread reread (libc) (hand) read write
--------- ------------- ---- ---- ---- ------ ------ ------ ------ ---- -----
ds20 NetBSD 1.4U 41 45 49 190 406 311 215 406 372
ds20-1 NetBSD 1.4U 148 148 46 183 393 312 236 393 371
ds20-2 NetBSD 1.4U 184 181 -1 181 397 199 244 396 371
ds20 NetBSD 1.4X 36 37 47 182 397 314 241 397 371
alpha-dec OSF1 T5.0 411 366 -1 438 1158 527 293 1153 486
Memory latencies in nanoseconds - smaller is better
(WARNING - may not be correct, check graphs)
---------------------------------------------------
Host OS Mhz L1 $ L2 $ Main mem Guesses
--------- ------------- --- ---- ---- -------- -------
ds20 NetBSD 1.4U 498 5 30 170
ds20-1 NetBSD 1.4U 498 5 30 171
ds20-2 NetBSD 1.4U 498 5 30 173
ds20 NetBSD 1.4X 498 5 30 172
alpha-dec OSF1 T5.0 498 5 30 184
A big gain for mmap latency, but no much changes otherwise.
I'm not sure what's wrong mith the local Communication bandwidths; things
were better with the non-optimised kernel. I'm going to re-run lmben
single-user.
I'll also try more agressive compiler flags.
--
Manuel Bouyer, LIP6, Universite Paris VI. Manuel.Bouyer@lip6.fr
--