home media network question

Have a question or want to start a discussion? Post it! No Registration Necessary.  Now with pictures!

My home network is a Windows workgroup with two Netgear GS108 rev 3 gigabit
unmanaged switches that can handle 9000 byte giant frames. Currently, three
machines on the network link at gigabit speed all running Windows XP:

1. The media server, a Tyan 2460 based dual Athlon 1.2 MP with 3 gigabytes
of DDR ECC RAM, a 64 bit 33 mHz bus, 3Ware 8 port 7400 RAID controller with
64 bit bus interface,  a 3 strip nonredundant 360 gig array, as well as
other JBOD drives. The ethernet controller is an Intel Pro 1000 MT 64 bit
dual gigabit port server adapter.

2. A desktop using the Via P4X266 chipset, 2.6 gHz P4, 2 gigs of DDR ECC
ram; 32 bit PCI at 33 mHz,  a Promise ATA100 hard disk controller, and a Pro
1000 MT desktop adapter.

3. The newest machine, ironically built for digital piano; a Gigabyte PCI
Express socket 939 design using the NForce 4 chipset, onboard gigabit
Ethernet, 4 gigabytes of dual channel 400 mHz DDR, a single 80 gig Seagate
ATA100 disk on the built-in controller.

Using the latest (August) version of the Intel drivers, which actually seem
a little worse than the March release, this is what I get for large files,
using the standard 1500 byte frame size:

Server --> machine 3:   130 mb/sec
machine 3 -> Server:    160 mb/sec
Server --> machine 2: 130 mb/sec
machine 2 --> server:   90 mb/sec (was 130 with the March drivers)

CPU usage on the server is about 20% or less.
CPU usage on machine 2 is around 40%
CPU usage on machine 3 is around 20-30%

Machine 2, in fact, is no better on the uplink than one of my Centrino
laptops, which manages 85 mb/sec up/down over a 100 megabit link via the
built-in Realtek controller, with CPU usage at 75% !

Giant frames set to 9000 result in about a 7% speed loss, but uplink on
machine 2 drops to near zero.
Buffer parameters for the Intel cards have been adjusted to the max, 2048
receive, 2048 send.
Disks are in a defragmented state.
These measurements were taken without other network activity.

Your comments on the likely bottlenecks are appreciated.

Site Timeline