Do not fear. David has the latest numbers, and has actually compared
them to SunOS 4.1.3 on the SAME hardware as linux.
Linux TCP throughput was 50% BETTER than SunOS. Admittedly TCP latency
was about 7% worse (lmbench reported 1530 usec latencies for SunOS and
1630 usec for Linux), and I'd like to fix that 7% too, but it's nowhere
near the numbers that people got with 1.2.8..
People should definitely take a look by the Stanford paper, but at the
same time keep in mind that TCP numbers have gotten 2-3 times better
since that paper, and NFS client read performance is up to 30 times
better depending on what you do (if it's not cached, it's not going to
be better, but...)
> I should point out that I obviously like
> linux, yet I do believe that it is productive to look at this sort of
> stuff and see what is/isn't wrong with it.
Linux has gotten a lot better in the networking performance department
especially in the 1.3.6x kernels (I've been working a lot on it, and so
has Alan and company). It's not just TCP, it's NFS too. I asked Mary
Baker to check out the NFS performance with the NFS cache, and linux was
faster than FreeBSD on NFS too (which traditionally has been a very very
weak point of Linux).
Admittedly FreeBSD was slightly better at writes and at "stat" calls,
but Mary and her team concluded it was a toss-up which was faster
overall, so things certainly have changed...
(The linux NFS client side still does not do any write coalescing, so
programs doing lots of small writes over NFS will still be slow. It's
being thought about..)
> Believe me, I would like
> nothing better than to see Linux beat the crap out of all the other
> OSes in every way possible (If only to annoy the BSD centrists ;). Keep the
> flames to private email please.
No flames, I would _really_ like you all to test out the latest 1.3.6x
kernels. 1.3.68 was good, and 1.3.69 is even better (it's a pre-release
of 1.3.69 that David was using for testing on the sparc). Things really
have changed..
Note that David hasn't compared against Solaris yet (so we've been
comparing the very latest linux release against the old SunOS kernel -
not what you could call a really fair comparison), so you should not
take this mail to mean that we kick ass.
What you _should_ take this mail to mean is that Linux as of the latest
kernels (and thus 2.0) will _not_ look particularly bad in networking
performance (SunOS 4.1.3 may be old, but Sun was known for good
networking so we're not trying to find a particularly bad system to
compare against to make our numbers look better).
And don't take my word for it - get 1.3.69 yourself and be a good tester
at the same time (ugh - I already got a report that it has some problems
compiling the ncp filesystem, but hopefully that small buglet is the
worst one).
Linus