Subject: Re: netperf vs 'real world' disk-net-disk throughput, BSD vs Linux
To: None <current-users@NetBSD.org>
From: Matthias Scheler <tron@zhadum.de>
List: current-users
Date: 09/11/2004 16:09:25
In article <20040910130316.148277a1@garlic.apnic.net>,
	George Michaelson <ggm@apnic.net> writes:
> FYAmusement, I've been testing a NetBSD, FreeBSD 5.3, debian/redhat combo
> with the same hardware with some people here, and while we can reliably
> get 400-600mbit/sec on back-to-back gig-e copper from netperf, when we do
> scp <bigfile> even to /dev/null, we're getting much much lower total
> throughput, around 100mbit max.

1.) "scp" uses encryption.

2.) "scp" is a userland application. So your data will go from kernel to
    userland and userland to kernel on each side which causes a lot
    of overhead.

Try NFS (which is kernel based) and "cp" (which cleverly avoids the copy of
the the data to userland) to get better numbers. My two P4s can move large
files at about 40MB/Sec from disk to disk which is quite close to the
maximum disk speed of about 50MB/Sec.

	Kind regards

-- 
Matthias Scheler                                  http://scheler.de/~matthias/