Date: Mon, 01 Feb 1999 05:06:59 -0800 From: David Greenman <dg@root.com> To: Greg Black <gjb@comkey.com.au> Cc: Greg Lehey <grog@lemis.com>, FreeBSD Questions <questions@FreeBSD.ORG> Subject: Re: Excessive collisions on Ethernet Message-ID: <199902011307.FAA02115@implode.root.com> In-Reply-To: Your message of "Mon, 01 Feb 1999 22:37:40 %2B1000." <19990201123740.27312.qmail@alpha.comkey.com.au>
next in thread | previous in thread | raw e-mail | index | archive | help
Anything less than 50% collision rate is okay and doesn't reduce the throughput significantly. All of the numbers below are in the noise. -DG David Greenman Co-founder/Principal Architect, The FreeBSD Project >> >> In the last few days I've noticed a really high number of collisions >> >> on my Ethernet. There are only 5 machines on the network, 3 of which >> >> are barely active, yet I see: >> >> >> >> (allegro, running 2.2.6-STABLE) >> >> ed0 1500 <Link> 00.00.c0.44.a5.68 43729816 45 43861788 12 977828 >> >> ed0 1500 widecast allegro 43729816 45 43861788 12 977828 >> > >> > That's only 1.1%. >> >> That's high for a small network. > >Well, that's not my understanding. I'd have considered anything >from 0.1% to 2.0% to be reasonable. > >> >> (freebie, running 4.0-CURRENT) >> >> Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >> >> ed2 1500 <Link> 00.80.48.e6.a0.61 11976144 12 13389307 0 988340 >> >> ed2 1500 widecast freebie 11976144 12 13389307 0 988340 >> > >> > And that's only 3.9% -- it's worse than allegro, but not by a >> > significant margin for the relatively small amount of traffic. >> > On my Ethernet, I have machines that report 0.01%, 9.6%, 0.7%, >> > 0.04% and 0.4%. The outlier is a machine that has been up a few >> > hours and was used for a large file transfer which blew its >> > average out of the water -- it'll be back to about 1% as time >> > goes by. >> >> Then you have problems too. > >I don't perceive any problems. And I'm hard pressed to believe >that an Ethernet that runs at pretty much its rated speed and >has collisions below 1% for all the significant data has >problems -- but maybe that's a matter of different opinions. > >> > I just did some 10 to 15 MB file transfers with NFS on my LAN (I >> > don't have FTP set up), and saw collision rates ranging from 25% >> > to 35% for those periods. I'd say that was pretty normal, given >> > the way that Ethernet works. >> >> No, that's way too high. Genuine collisions happen on an Ethernet >> when two systems want to send a packet within a very small time >> window: first they look and listen, and if nothing is there, they >> send. If two send at pretty much *exactly* the same time, there will >> be a collision, from which they recover. > >Surely an ftp transfer involves plenty of packets going in >*both* directions and so can be expected to produce a fair >number of collisions? > >> If only one machine is sending at a time, there won't be collisions. >> If only a few are sending, there won't be many collisions. In >> general, you can consider 1% collisions to be an acceptable number. > >I reported 1% collisions above, but you said I had problems. >I'm not sure what level of collisions you feel represents a >problem. As I understand the way it works, errors are a problem >but collisions are ok in some (smallish) numbers. > >> There's another thing of interest in this picture: in the example I >> showed above, collisions were at 40%. At this level, traffic on an >> Ethernet is becoming highly congested. Yet I got a transfer rate of >> just under 1 MB/s for the transfer, which suggests to me that the >> statistics may be bogus. A thing that just occurs to me is that it's >> always the ed driver that reports so many collisions, whereas the >> others don't. What kind of Ethernet board are you using? > >Here are some data from my Ethernet: > >maxim up 73+01:12, 0 users, load 0.05, 0.07, 0.07 >alice up 30+06:32, 2 users, load 0.13, 0.09, 0.08 >alpha up 4+03:55, 3 users, load 0.13, 0.09, 0.03 >bravo up 4+03:55, 0 users, load 0.00, 0.00, 0.00 > >Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >ef0 1500 <link1> 00:60:08:ac:d7:1b 5781293 16 11978534 12 7189 >ef0 1500 203.9.155.2 maxim 5781293 16 11978534 12 7189 >ef0 1500 192.168.1 192.168.1.1 5781293 16 11978534 12 7189 > >Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >eb0 1500 <link1> 00:60:08:a8:17:5d 4898485 1 2485013 0 52390 >eb0 1500 203.9.155.2 alice 4898485 1 2485013 0 52390 >eb0 1500 192.168.1 192.168.1.2 4898485 1 2485013 0 52390 > >Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >xl0 1500 <Link> 00.10.4b.18.1f.b9 25057 0 26207 0 4340 >xl0 1500 192.168.1 alpha 25057 0 26207 0 4340 > >Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >xl0 1500 <Link> 00.10.4b.18.1f.a8 15808 0 8428 0 2 >xl0 1500 192.168.1 bravo 15808 0 8428 0 2 > >maxim is an old 486-33 running BSD/OS-3.1 with a 3C509, alice is >a Pentium-166 running BSD/OS-3.1 with a 3C900, alpha and bravo >are Celeron-300A's running FreeBSD-2.2.8-Release with 3C900B's. >The errors on maxim were from the time when it booted and the >rest of the LAN was out to lunch. The others all booted more >recently on to a live network. The experiments I tried with the >NFS copies were between alpha and alice and bumped alpha's >collisions from virtually zero to the present figure. > >> > The more interesting statistics are the overall figures over time, >> > and the ones you give at the start aren't too bad. Yes, freebie is >> > a bit high -- but 4% compared with 1% is not a big factor, >> > especially with the fairly low total amount of traffic. I'd watch >> > it for a bit longer before deciding it was a real problem. After >> > all, if you're getting 1 MB/s on a 10 Mb/s LAN, you're doing pretty >> > well :-) >> >> Well, the values for freebie are now: >> >> Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll >> ed2 1500 <Link> 00.80.48.e6.a0.61 16493376 34 21501269 0 2326175 >> ed2 1500 widecast freebie 16493376 34 21501269 0 2326175 >> >> Since yesterday, that's 8 million more output packets and 1.3 million >> collisions. That suggests that something might be getting worse. > >Yes, it does look as though it's getting worse. You may indeed >have a fault somewhere. I'd certainly check the cables first, >and try swapping some of them around before pulling cards from >machines. (I have to pull video cards from five machines >tomorrow to swap them for something that works with FreeBSD, so >I'm a bit jaundiced about pulling cards at the moment :-) ) > >Greg > >-- >Greg Black <gjb@acm.org> > > >To Unsubscribe: send mail to majordomo@FreeBSD.org >with "unsubscribe freebsd-questions" in the body of the message To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-questions" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199902011307.FAA02115>