Gigabit Network Adapters on Platform TYAN Trinity GC-SL.
Part One: 32bit PCI Interface

Part 1.3: Behaviour on dynamically changed blocks of transmitted data (NetPIPE), average retail prices, conclusions

In this case the graphs show the adapter's stability (speed invariability) in the whole range between minimal and maximal 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.

Bandwidth dependence on the size of the transmitted packet in each card, NetPIPE, Linux

No deviations from the norm were found in 3Com 3C996B-T.

We could only test CNet ProG2000L with Jumbo frames disabled, but even this test shows there's something wrong with the adapter (or its driver). A number of packet sizes are marked by performance falls.

D-Link DGE-510T is doing all right.

Hardlink HA-64G is all right too.

The TRENDnet TEG-PCISXM2 optical card based on Marvell Yukon 88E8010 has performance falls at MTU (and consequently, Jumbo frame) sizes up to 6000, while it's all right at 9000. As will be seen below, the picture with the SysKonnect adapter (based on the same chip) is identical.

Everything is OK with ZyXEL Omni Lan PCI G1.

The SysKonnect adapters were gathered into one group as they all have identical results: falls at some sizes up to MTU=6000, no falls at 9000, the larger the packet (and the smaller the Jumbo frame), the more probable a performance fall. And because it repeats with all the adapters (on different chips) we'd suggest that the driver is to blame for this.

Unfortunately, this anomaly overshadows good PI readings of SysKonnect adapters.

Bandwidth dependence on the size of the transmitted packet for each card (the data are grouped according to the MTU size).
NetPIPE, Linux

For those interested in this aspect, here are the dependence graphs with no comments added.

The graphs have a lot in common with the adapters' peak bandwidth in NetPIPE and the speeds at various sizes of transmitted data blocks.

Prices

Apart from PIs, the prices, too, play an important role in the choice of the adapter. All the adapters mentioned in the article are available in Moscow, except the SysKonnect boards (at least, there was no information about them on price.ru).

Conclusions

It is difficult to draw conclusions in roundups, especially if there is no undisputed leader. 3Com shows good performance and doesn't load the CPU much (especially in Linux). But the price makes all the difference... Hardlink shows moderate performance and is fairly cheap. The SysKonnect and Trendnet adapters have high performance and a moderate CPU load, and besides, they support a great number of OSs. But performance falls at some packet sizes in NetPIPE prevent them from becoming evident leaders. Perhaps, the new driver version will change the situation. Despite the fact that a lot of users integrate Realtek 8169 into their motherboards (thus getting a free gigabit onboard), it didn't show its best side on Tyan.

What would you prefer? It all depends on what you need. Anyway, this review aims at comparing just the adapters most of which are integrated into motherboards. Because indeed, it's unwise to install a high-speed gigabit adapter on a 32-bit PCI bus, as its potential can't reveal on it.

The second part of the article will deal with testing the same adapters on a 64-bit PCI bus. That's when we'll be able to make real conclusions.

Navigation:

• part 1.1: general description and testing technique;
• part 1.2: performance, CPU load, and PI;
  part 1.3: behaviour on dynamically changed blocks of transmitted data (NetPIPE), average retail prices, conclusions.