Subject: Re: [netatalk-admins] getting a new nic
From: Richard Knuckey (richardk@iprolink.co.nz)
Date: Wed Mar 11 1998 - 15:55:11 EST
At 8:09 am +1300 12/3/98, Eric KieslingThompson wrote:
>I confess that I built a server using an ne2000 clone. When things get busy
>at the firm, with big graphics apps running, and big file copies, the
>network gets REALLY slow. I've upgraded the RAM in the server to 96MB (will
This won't help much with netatalk. Netatalk uses very little RAM. Some
production servers (10baseT) I've build have only 16MB, and achieve >1MB
Sec transfer rates with asun's TCP netatalk.
>be 128 when I get the next chip), I'm going to add a hub (two already, but
>pretty full), and I'm going to put the server on a shorter length of cable
>(about 50' right now, but shouldn't even matter under 300', right?). I'm
The length of ethernet cable does not affect speed in any way. What will
increase overall network bandwidth is a switched hub, but only if your
traffic patterns should alot of inter-machine traffic. It 90% of your
traffic is between the macs and the server you will not see any speed
improvement, unless your switching hub as a 100BaseT port, and the server
is on 100BaseT ethernet.
>running Slackware 2.0.30, and netatalk 1.4b2. I would like to use asun's
>stuff, but no luck so far with compiling (still working through 6,000,000
>emails :) ).
2.0.30 has major bugs in memory management in the kernel, which affect high
seped networking and disk access sometiong chronic. I would recommend
upgrading the kernel to 2.0.32.
>BUT, I have a feeling that the ne2000 clone may be a major factor. I'm
>going to get a new nic, and I'm wondering if any of you have a
>recommendation.
The quality of NE2000 Clones varies wildly, but I've not had any problems
as long as the card works at all. As for alternatives I can highly
recommend any recent Digital Chipset card, using either the DE45x or Tulip
drivers (depending on manufacturer)
>The server is an AMD K5-PR100, with isa and pci. The network is 10MB
>ethernet. So far, all clients are MACS, but I'm expecting to add some PC's.
The Number 1 biggest speed killer is IDE hard drives. Even with Intel
Triton chipset bus mastering, 50% of processor time is taken up in
transfers to and from the hard drive. Add that to the overhead in
performing checksuming on your Appletalk packets (the checksuming routines
for tcp in the kernel are faster than the appletalk ones), and you end up
with slower performance that you'd want. I would expect you'd be getting
around 500MB/sec on this machine. Upgrading to asun and using tcp should
increase this to arond 900M/sec.
If you use SCSI hard drives and a controler like an Advansys, or Adaptec
2940 (of its many variants) etc then you should except around 650M/sec on
appletalk and 1050MB sec on tcp. Processor overhead for a SCSI card is
usualy in the 6% region.
The Digital Prioris server I am currently using achieves the maximum
theoretical rate of 1250MB/sec as verified by Helios LanTest 2. This has an
on-board Adaptec Ultra-Wide dual channel adapter talking to ultra-wide hard
drives (hdparm gives 9.7MB/sec from the hard drives)
As for 100BaseT networking, you will need more RAM as this requires larger
buffers to operate efficiently (32MB should do it, unless the machine has
other tasks to perform), you require SCSI, and fast hard drives, and a PCI
100BaseT Ethernet card. Avoid Intel EtherExpress Pro 100's. They do not
work. I would recommend a Digital chipset based card (D-Link, SMC, Digital
etc). You also need a pretty decent motherboard with good throughput if
your going to attempt to use the machine for anything else, the throughput
on most PC motherboards is around the 35MB/Sec mark, when you are trying to
push data around at 20MB/sec (10MB from HD->RAM, 10MB from RAM->Ethernet)
A server in a pre-press house I set up, with an Advansys SCSI card (SCSI-II
Fast, not Ultra), 6 4GB Hard drives, 100BaseT Digital ethernet card and
P200. HD transfer speeds are 7MB sec, and ethernet transfer speeds are
5.5MB sec (verified by Helios Lantest).
When this was put in place, 2.0.30 was the current kernel release and we
had problems getting raid working, and out of memory errors when trying to
format drives larger than 4GB, plus most of the fast ethernet cards we
tried crashed with buffer overflows after a while, therefore we were stuck
with what worked, which was a slower system than what we originaly
anticipated. These problems have been fixed in 2.0.32.
With raid 0 over two SCSI controlers, which should increase the hard driver
transfer rates to around 14MB/sec. I would expect to be able to saturate
100BaseT ethernet with this machine.
This archive was generated by hypermail 2b28 : Sat Dec 18 1999 - 16:31:29 EST