Subject: Re: file transfer times
From: Christopher E. Brown (cbrown@woods.net)
Date: Tue Apr 02 2002 - 12:37:17 AKST
On Tue, 2 Apr 2002, Peter Q. Olsson wrote:
> Hello all-
>
> I have a lan sitting behind a firewall that consists of several linux boxen and
> a Sun ultra2 running solaris. I have found a great disparity between file
> transfer rates among various boxes.
>
> For example, transfering the same 1.2 MB file btwn box A and box B:
>
> going in the direction A -> B I get a transfer rate of ~2000 kB/s
> going in the opposite direction, B->A, I get a transfer rate of ~50 kB/s
>
> This is quite a repeatable result, though of course the numbers vary somewhat
> each time. Still the order of magnitude difference is the same.
>
> Other combinations of boxes show intermediate results ~200-400 kB/s either
> direction. scp also shows similar results, about a factor of 10 btwn directions
>
> Any ideas as to what gives here?
>
> PQO
Please supply more data, any answer to the above would be nothing more
than wild guessing.
Machine types, kernel versions, and ether card
CPU speed and Solaris version for the Ultra
If is a HUB or a Switch, and what type/speed
EXAMPLE
PII-300, Ne2000, Linux 2.2.18
P-4 1600, Intel EtherExpress 10/100+, Linux 2.4.14
Ultra2 300Mhz, SunFE (offboard) 10/100, Solaris 8 (current patches)
Kingston KNE8TP/W 10Mbit HUB
And a description
Transfers FROM Solaris 8 machine TO PII-300 SLOW, copy command issued
ON Solaris machine (or) copy command issued on Linux machine. (push
vs pull matter for measurement accuracy in many applications (ftp and
scp included)).
It could be anything from IRQ issues, there are a few interactions
with TCP windows and Solaris that come to mind, to duplex issues with
a switch.
There are also issues with transfer stats reporting on files that
small. Most transfer readings are approximate, and can yield strange
results on averaged rate with files small in comparison to the
transmit buffers in a machine. (The application normally tracks how
much data the kernel has accepted, *not* how much the remote has
accepted, a 1.2MB file and a 256k - 1024k transmit buffer + a large
TCP window can inflate rate readings. This is why most benchmarking
is done with files > 300MB).
It *sounds* line a 100Mbit network (1200Kbyte/sec is topping a
10Mbit), but if so the 300 - 400 Kbyte/Sec rates are pretty low. If
it is a 10Mbit network than 2000Kbyte/sec reading is flawed. If is
switched mixed 10/100 network than things are even more interesting.
-- I route, therefore you are.
This archive was generated by hypermail 2a23 : Tue Apr 02 2002 - 12:39:43 AKST