Date: Mon, 18 Nov 1996 08:54:57 -0600 (CST) From: Joe Greco <jgreco@brasil.moneng.mei.com> To: alk@Think.COM (Tony Kimball) Cc: isp@freebsd.org Subject: Re: changed to: Frac T3? Message-ID: <199611181454.IAA02266@brasil.moneng.mei.com> In-Reply-To: <199611180009.SAA01198@compound.Think.COM> from "Tony Kimball" at Nov 17, 96 06:09:27 pm
next in thread | previous in thread | raw e-mail | index | archive | help
> Quoth Joe Greco on Fri, 15 November:
> : I was reliably routing 5000pps the other day on a Pentium 100...
> : and it did not seem particularly stressed out.
>
> Has anyone constructed a performance model for routing on x86
> hardware? You only need to improve this performance by a factor
> of 3.4 in order to attain Dennis' magic 17 kpps. That does not
> seem very far out.
*cough* 9000pps. Only involving two de interfaces on Trantor...
% netstat -I ed0 -b 10 | cut -c61-120
input (Total) output
packets errs bytes packets errs bytes colls
1692 0 496741 1692 0 491467 9
60613 0 4082655 58893 0 3415430 22236
90373 0 6071290 88007 0 5102437 33898
91345 0 6058466 90056 0 5146746 32342
91830 0 6109491 89431 0 5122133 33697
91946 0 6020587 90123 0 5034919 35137
97586 0 6361299 94241 0 5224866 35690
89598 0 5939400 88427 0 5037640 33711
I am seeing exponential CPU idle drop off though, it is running 20-40%
idle.
We are pumping! Ok let's go for the big finish... let's hit
10Kpps if we can... unfortunately I have to involve ISA crud
machines to do this....
YES!!! This rules :-)
% netstat -I ed0 -b 10 | cut -c61-120
input (Total) output
packets errs bytes packets errs bytes colls
115709419 5991 1515971169 115823590 468 1410765988 391521
44890 0 3126746 44894 0 2776626 786
46338 0 3210192 46344 0 2849657 898
80950 0 5218615 70815 0 3826148 11247
105120 0 6794313 90143 0 4804222 18815
89825 0 5899932 79135 0 4345818 17300
113552 0 7247128 99339 0 5181548 22321
91177 0 5939665 80887 0 4338463 17897
113245 0 7358616 97373 0 5205285 22899
113625 0 7314844 95726 0 5054561 23281
118345 0 7654968 100091 0 5290368 22678
109026 0 7093593 92468 0 4966475 20996
127815 0 8175061 102896 0 5319967 28121
I am definitely running out of CPU cycles though.. 10% idle now.
procs memory page disks faults cpu
r b w avm fre flt re pi po fr sr f0 w0 in sy cs us sy id
0 0 0 31544 2860 1 0 0 0 0 0 0 0 23513 38 8 1 96 4
0 0 0 31544 2860 1 0 0 0 0 0 0 0 24700 32 7 0 88 12
1 0 0 31544 2860 1 0 0 0 0 0 0 0 23560 37 7 0 97 3
0 0 0 31544 2860 1 0 0 0 0 0 0 0 35945 47 10 0 95 5
0 0 0 31544 2860 1 0 0 0 0 0 0 0 26330 30 8 0 98 2
Conditions of the test...
The packets in question were teeny UDP datagrams mainly because the
collision rate grew excessively large if they were the original 1K or 256
byte packets I was playing with. The packets were blasted at an arbitrary
host on one de interface, from host(s) on the other de interface(s).
Note that production network traffic was still traversing this machine
during these tests, and while I noticed a very moderate slowdown in my
telnet traffic, it was not bad at all. Also, since most of the production
traffic traverses the NE2000 interface, there is some built in performance
degradation... I am thinking maybe 12000pps on a P100 is not beyond
achievability. However, it would take someone with 100mbps Ethernet
cards to try it with reasonable packet sizes.
I like that interrupt rate though :-)
... JG
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199611181454.IAA02266>