On 2000-12-14, "Fulvio Risso" <[EMAIL PROTECTED]> wrote: > Buffer sizes > We did not make any test about creating 1MB buffers. However our > architecture does not have the problem that "large buffer" = "large > time used to transfer this buffer to user level" because we are able to [snip] > Context switches > Are kept the lowest, because several packets can be transferred at the > same time. Erm... but if you've not enlarged the buffer from the default, you've not kept context switches "at the lowest." Sure, there isn't one switch per packet. But there could be far more packets per switch if you'd enlarged the buffer (note there's some diminishing returns as you exceed different cachesize boundaries, i.e. performance will actually drop when you working set gets too big for L2 cache, etc). > UFS filesystem > We used FAT to strore packets because the UFS filesystem was on a > second hard disk, so that the disk could have made some difference. We > can reinstall BSD and repeat the test. We'll do that for sure. I suspect this will dramatically change things. Helping a friend who was working on 100mbit sustained capture on FreeBSD, I realized he was writing to a FAT filesystem (on removable drives) for ease of portability. I suggested he switch to UFS; he did so and realized a 200% performance improvement (3x). This was before softupdates, which should make things even better. FAT is an awful filesystem to begin with, and I'd not be surprised if its implementation in FreeBSD isn't as good as it could be (because, why bother?). I'm afraid, this one simple fact completely distorts your results. Of course, even if you switch to UFS it won't quite be apples-to-apples. But at least, you will be using the native/best/most appropriate filesystem for each platform you're testing. Right now, whenever you start writing packets to disk under FreeBSD, you're testing how much FAT blows, and not much else. -- Hank Leininger <[EMAIL PROTECTED]> - This is the TCPDUMP workers list. It is archived at http://www.tcpdump.org/lists/workers/index.html To unsubscribe use mailto:[EMAIL PROTECTED]?body=unsubscribe
