Testing Gigabit Network Adapters on platform TYAN Trinity GC-SL.
Part Two: 64bit PCI Interface

Part 2.3: adapters' performance with dynamically changed data packets (NetPIPE), average retail prices, conclusions

In this case the graphs show the adapter's stability (speed invariability) in the whole range from minimum to maximum data blocks. If the graph line is smooth, then it's all right, gaps indicate that the adapter has a performance problem at this particular block size.

Throughput dependence on the transmitted packet size for each adapter, NetPIPE, Linux

Both embedded gigabit chips from Intel demonstrate somewhat "twitched" graphs, though all the gaps are insignificant.

No anomalies were detected in 3Com 3C996B-T.

We managed to test CNet ProG2000L only with giga-frames disabled, but even in this case there is something wrong with this adapter (or its drivers). It shows performance falls with some packet sizes.

Hardlink HA-64G is also doing all right.

TRENDnet TEG-PCISXM2, based on a Marvell Yukon 88E8010 chip, shows performance falls with Jumbo frames below 9000 bytes. The smaller is MTU size, the more gaps appear. The situation abruptly changes only with 9 KB frames — gaps disappear.

ZyXEL Omni Lan PCI G1 is OK.

I have united all adapters from SysKonnect in one group, because they all demonstrate the same gaps in performance graphs, similar to TRENDnet TEG-PCISXM2. And again like in TEG-PCISXM2, the gaps disappear with 9 KB frames. As all these adapters (SysKonnect and TRENDnet) were tested using the same driver, it's most likely the driver's fault.

That's why, though adapters from SysKonect demonstrated excellent performance, the bug with speed drops with certain packet sizes is irritating.

Throughput dependence on the transmitted packet size for each adapter, grouped by MTU.
NetPIPE, Linux

For interested readers we publish uncommented graphs of throughput dependence on transmitted block size in NetPIPE, where the data is grouped by the MTU size. These graphs allow to compare the real-time performance of all tested adapters with a certain MTU size.

Availability

Besides performance and CPU load, any adapter is characterized by its price. It is not taken into consideration in the index formula, it is just displayed below in the form of a table. Unfortunately, this table does not include cards from Intel, as we tested only embedded solutions. All the adapters reviewed can be bought in Moscow, except for the cards from SysKonnect, at least we didn't manage to find them in price.ru.

Average retail price of gigabit adapters for the present moment:

Conclusions

• Our tests demonstrated that next time it will make sense to separate server and desktop adapter versions.
• Repeating situation with a relatively low adapter speed in Windows 2000 requires testing adapters in Windows 2003 Server as well.
• CSA interface allows adapters (at least adapters from Intel) to demonstrate very high performance approaching the performance of adapters on a 64bit PCI.
• Test leaders are adapters from 3Com and "at a stretch" from Intel (because of not so high speeds in Linux). Right after them is a tight group of cards from SysKonnect.
• Not all new adapters have good drivers, take for example cards from CNET and Zyxel.

Navigation:

• Part 2.1: overview of adapters and test methods
  
• Part 2.2: adapters' performance, CPU load, and PI
  
• Part 2.3: adapters' performance with dynamically changed data packets (NetPIPE), average retail prices, conclusions