NETGEAR GA620 - PCI Fiber Card Manual

White Paper on GA620/GA620T Gigabit Ethernet Card
Performance
Gigabit Ethernet adapters represent the next step in the evolution of
network performance. The widespread adoption of Fast Ethernet at the desktop
has created a bottleneck at the server, specifically at the Server adapter. Some
companies have created Server adapters with multiple 100 Mbps connections,
while others have populated several PCI slots with 100 Mbps adapters to boost
Server bandwidth beyond Fast Ethernet. Gigabit Ethernet delivers performance
beyond these aggregation band-aids, by delivering speeds approaching 1 Gbps.
Network Configurations
Due to the fact that Servers benefit most from Gigabit speeds, the adapter
card drivers are optimized for that environment. In most networks, the server
distributes information to multiple clients. For this reason, the driver has been
optimized to perform best in a one-to-many environment. Using the test program

Chariot running against twenty-five 100 Mbps clients, Gigabit speeds are
consistently reproduced in lab.
Test results will vary depending on a number of factors including the type
of test, the speed and performance of the computers being used, and the
network topology. To measure the capabilities of the adapter, the test should
stress the network and not the other components of your system. Programs
such as Ganymede’s Chariot, and Microsoft’s NTTTCP accomplish this goal
quite well. These tools provide a multi-threaded, asynchronous performance
benchmark for measuring achievable data transfer rate.
Back-to-Back Test
One test that does not accurately measure adapter performance is a back-
to-back test. The driver is optimized for one-to-many scenarios, so testing in a
back-to-back environment will yield slower results. More than 500 Mbps
throughput is still possible using a multi-threaded testing program for testing
network performance, such as Microsoft’s NTTTCP program. While one is
unlikely to see the full capabilities of a Gigabit Ethernet card when tested back-
to-back, it is still possible to get a feel for the capabilities of the adapter.
File Transfer
Another test that does not accurately measure adapter capability is a file
transfer. Copying a file from one computer to another tests a hard drive’s
performance rather than network performance. In a test like this, the limiting
factor is the performance of the computers, rather than the performance of the
network. The copying process adds the overhead of the disc drives, as well as
the operating system handling the copy. The test results indicate the computers’
ability to process these additional requirements rather than the abilities of the
connection between them. In a test like this it is possible to see results less than
20 Mbps on any card, which is slow for both a 100 Mbps card and the 1000 Mbps
card. Surprisingly, the Gigabit adapter will probably be the slower of the two
cards in this test.
The reason a 1000 Mbps adapter would perform below a 100 Mbps
adapter is quite simple in this test. In order to achieve Gigabit performance,
Gigabit adapter drivers contain additional intelligence beyond Fast Ethernet
drivers. This intelligence creates more overhead than a standard 10/100 card.
When stressing the network at higher speeds, this overhead is negligible, and
improves to the performance of the network. When the cards are run at minimal
performance, such the file copy test, the Gigabit card will have slightly more
overhead than a standard 10/100 card, causing slower performance. However,
as more stress is placed on the network, such as video or multiple system file
transfers, the gigabit card will start to outperform a 10/100 card. Ultimately, the
Gigabit adapter can deliver 10 times the network throughput of a Fast Ethernet
adapter.

Testing Results
1. Using Ganymede’s (or netIQ) Chariot
One-to-Many Traffic Flow
OS environments Throughput (Mbps)
Windows NT4.0 server to 25 pair clients 915.775
Windows 2000 server to 25 pair clients 529.452
2. Using Microsoft NTTTCP
Back-to-Back Traffic Flow (Windows 2000 to Windows 2000)
Sessions / Threads Throughput (Mbps)
30 594.843
16 576.568
8509.942
Back-to-Back Traffic Flow (Windows NT4.0 to Windows NT4.0)
Sessions / Threads Throughput (Mbps)
30 562.813
16 552.356
8555.565
Test Configuration (10/25/2000):
1. One-to-Many:
This configuration was comprised of one chassis switch with Fast Ethernet
connections to 25 clients (18 Windows NT Workstation, 4 Windows 2000
Professional and 3 Windows 98) and 1-gigabit uplink to the server (Dell
PowerEdge 4400/600 server running either Microsoft Windows NT 4.0 Server or
Windows 2000 Advanced Server).
2. Back-to-Back:
The back-to-back test bed configuration was comprised of one Dell PowerEdge
4400/600 server and one IBM Netfinity 6000, both runningeither Microsoft
Windows NT 4.0 Server or Windows 2000 Server.

The latest network interface cards and drivers were used in all tests. Throughout
the testing used two different server machines:
1. IBM Netfinity 6000 Server (600MHz, 2GB RAM, BIOS version: 1.02) with
dual boot between Microsoft Windows 2000 Server and Microsoft
Windows NT 4.0, SP6;
2. DELL PowerEdge 4400/600 Server (600MHz, 512MB RAM, BIOS version:
A04) with dual boot between Microsoft Windows 2000 Server and
Microsoft Windows NT 4.0, SP6.