Conclusion

Three general types of test were performed, providing information about three general uses. The Transfer Time tests included FTP and NFS, but are also indicative of the relative performance that would be seen through the use of HTTP, RCP and SCP. This activity would be used when files are sent around the network, or shared. This type of activity is likely to be the most useful of a GigaBit connection to a desktop workstation.

The Bandwidth gives useful information about how well this card could be used to stream information into a system, but not onto disk. The most common use of this activity would be remote visualization, where images are rendered remotely, then displayed back to a desktop workstation.

The Latency tests show a considerable amount of information, and they show that the times are usually related. Latency times potentially effect every network activity. Some network activities have data transfer patterns that are relatively immune to long Latencies. The Latency values are most important when activities are performed with a remote database. Other activities with small packets that require fast handshaking (i.e. MPI applications) will also be significantly effected by Latency times. Low Latency is usually less important on a Client system than on a Server.

The first conclusion is that if you are using a NetGear GA622TNA in a high performance server, replace it. This card would be OK (not great) to use in a client, or a low performance server, if you have an unused card laying around (it's better than 100Base-T). Even in this use, the NetGear GA622TNA is the worst of the tested cards. If possible, avoid this card. This conclusion is based on the high Latencies, low Bandwidths and long Transfer Times that this card exhibits in all tests, especially when used in the server. This is not the sort of bottleneck that should be placed in a server. The remaining conclusions assume that this card is not being used in the server.

Here are the overall averages of each of the three types of test. As before, Latency and Transfer Time are better when they are low, and Bandwidth is better when it's high.

Averages	3Com 3C996B-T	Intel 8390MT	NetGear GA622TNA	SMC 9552	SMC 9562	SysKonnect 9D21
Latency	147.9	148.0	511.8	103.2	101.9	133.3
Bandwidth	87.3	82.7	73.4	59.1	58.7	84.1
Transfer Time	57.7	67.3	76.8	70.1	69.6	59.9

The above information readily shows that the cards have different strength and weaknesses. As previously, the NetGear GA622TNA gives a very poor showing. The long Latencies of this card are the probable cause of the long transfer times. This is something that no amount of tuning will be able to fix.

The 3Com 3C996B-T has the highest Bandwidth and lowest transfer times. This would make it the best card for most desktop client applications.

The Intel 8390MT card has slightly worse performance that the 3Com card, but it costs about half as much. For many applications, this would be a reasonable tradeoff.

The SMC cards have the lowest Latencies, making them the best performer for MPI or Database applications. It is probable that increasing the size of the network buffers will generate a significant improvement in both the Bandwidth and transfer times.

The SysKonnect 9D21 card has been discontinued. It's performance is better than the Intel card, at a similar price (when it was available). If you can find this card, use it. The street price for this card was around $50, which is significantly lower than the 3Com card, while it's performance was only slightly worse.

I would place the cards in the following overall order, when used in a desktop class system.

1. 3Com 3C996B-T
2. SysKonnect 9D21 (discontinued)
3. Intel 8390MT
4. SMC 9552
5. SMC 9562
6. NetGear GA622TNA

The above ranking is based on what the tests showed when I ran them.

In addition to the above, my tests show that the fastest cards in a homogeneous server environment (64-bit PCI) are the SMC cards. This is important to those people who build clusters. It is important to realize that there are other high-end server cards that fall into the price category of this test, but were not tested. These other cards may well have higher performance.

I expect to see several changes in the short-term future:

1. I expect the prices of 32-bit GigaBit cards to continue to drop. This is based on the overall simplicity of the cards (relatively simple production), their high overall performance as compared to the 64-bit cards (high demand), and their current low price (high demand). Currently, 32-bit GigaBit cards are priced where 10/100 cards were two years ago.
2. I expect GigaBit Copper switches to decrease in price. This is due to an increasing demand for them, which should cause greater production.
3. I expect minimal price drops in Fiber cards (only copper cards were tested here). A significant amount of the cost of a Fiber GigaBit card is in the optics. With the increasing availability (and decreasing price) of Copper GigaBit cards, I expect that the Fiber GigaBit production will decrease, which may even cause the price to climb slightly.

Prev

Index

If you have any comments or suggestions, please E-mail webmaster@accs.com

© 2004 - Ashford Computer Consulting Service