Messages in this thread | | | Date | Fri, 16 Apr 1999 15:50:03 -0700 | From | David Miller <> | Subject | Re: bad lmbench numbers for mmap |
| |
Date: Sat, 17 Apr 1999 01:34:18 +0530 (IST) From: V Ganesh <ganesh@vxindia.veritas.com>
I tried lmbench 1.9 on an ultra5 (270 MHz, 128 MB RAM) running linux 2.2.2 and compared with an identical machine running solaris 2.6(except it had 64M). linux blew solaris away in most of the benchmarks except mmap latency.
...
so what's up with mmap ?
There are two things which changed in the mmap/munmap code paths recently, and I know the numbers were competitive when I was running lmbench a lot several months ago.
The two changes are:
1) Using VMA trees (I really don't think it accounts for the slowness being seen here)
2) The freeing up of page tables at munmap
I think #2 could be a possible culprit.
Linus, did you see any performance degration of lat_mmap when you added the free_pgtables() stuff into 2.2.x? I think such a problem would trigger if lat_mmap is doing large enough mmap's and they are in the right place.
Mr Ganesh, can you post here the full lmbench raw output for the lat_mmap test? It should be in the results/sparc64-linux/xxxx files lmbench made as it ran.
Later, David S. Miller davem@redhat.com
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |