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

Part 1.2: Performance, CPU load, and PI

The results are divided into two sections. The first one includes diagrams of adapters' average maximal speeds, CPU load at that moment, and performance index (PI). The second one comprises researches of adapters' activity when the size of data block in transmission is changed on different MTUs (using programs NetPIPE and Linux).

During the tests we took the reading of an adapter's average maximal bandwidth for a certain time span. And if the program allowed it, CPU load for that period was also read out. But taken separately, the readings of adapter bandwidth and CPU load are opposite in their values (who needs a gigabit speed if the CPU is loaded at a 100 percent?), that is why we came up with the idea to unite these two readings in one formula.

The formula of a card's performance seems evident: you just have to divide transmission speed by CPU load. But in reality, it turned out to be a bit more complicated. Let's take, for example, a card's bandwidth equal to 100 Mbits (on a gigabit adapter) and CPU load equal to 5 percent. The division procedure gives a very high index, but this value is far from reality. That made me and Alexey Kuznetsov introduce the notion of Relative Index (RI) and multiply the results by the ratio between the real and the peak transmission speeds.

Thus, PI can be found in the following way:

PI = T/C * T/1000, where

• T is the maximal speed in the test, Mbits
• C is CPU load in the test, percentage
• 1000 is the adapter's maximal theoretical speed, Mbits

NTTTCP results, Windows

I must confess that I was disappointed by the NTTTCP test in this particular case. Most results were worse than those received in other tests, and all attempts to improve them ended in failure. Next time, we'll probably do without this test.

adapters' maximal speed with different Jumbo frame sizes

CPU load with different Jumbo frame sizes during testing

adapters' PI with different Jumbo frame sizes, the most typical diagram in NTTTCP

Most adapters have close results. The low speed of TRENDnet TEG-PCITX2 (the old dual-chip model) and an extremely low speed of CNet ProG2000L on the RTL8169 chip are really disappointing. CNet ProG2000L is also marked by a very high CPU load, especially with Jumbo frames disabled. Of course, it is typical of any Realtek-based card, but in this case, the speed is very low too. D-Link DGE-510T has a very high CPU load with Jumbo frames disabled, as well.

However, all that we have mentioned is clearly seen from the comparative diagram of PIs. 3Com, not at all a leader on diasbled Jumbo frames, shows very good results at 6000 and 9000 bytes. Also noteworthy are stable and good results (close to each other) of SysKonnect adapters. The Realtek-based card is the outsider: its index is below zero.

Iperf results, Windows

Unfortunately, the program doesn't estimate CPU load, and we have no diagrams of the load and PI.

adapters' maximal speed with different Jumbo frame sizes, Iperf

The results are congruous to those of NTTTCP, only a little better. All conclusions concerning NTTTCP can be applied to Iperf as well.

NetIQ Chariot results, Windows

The test allows to take the readings of both adapter's bandwidth and CPU load at that moment, and therefore has all the three diagrams including one with the Performance Index.

adapters' maximal speed with different Jumbo frame sizes, NetIQ Chariot

CPU load with different Jumbo frame sizes during testing, NetIQ Chariot

adapters' PI with different Jumbo frame sizes, the most typical diagram in NetIQ Chariot

Once again, the results and conclusions are similar to the NTTTCP test. With Jumbo frames enabled, the 3Com adapter shows a somewhat higher speed than others but it also loads the CPU more. Strange as it may seem, when the frame size is 3000 bytes, the 3Com adapter demonstrates a very high speed with a relatively low CPU load and thus has the best PI. But at 6000 bytes and higher, its CPU load grows dramatically while its speed only shows an insignificant increase. In most tests, it results in the 3Com's loss to the SysKonnect adapters that go neck-and-neck. And they also lead at 1500 bytes (that is, with Jumbo frames disabled). Although the latter fact is not so interesting considering there are no motherboards for the moment that would have integrated gigabit controllers used in SysKonnect adapters.

NetIQ Chariot results, Linux

In the previous testing, Windows and Linux results differed greatly. Let's see what we've got this time.

adapters' maximal speed with different Jumbo frame sizes, NetIQ Chariot

CPU load with different Jumbo frame sizes during testing, NetIQ Chariot

adapters' PI with different Jumbo frame sizes, the most typical diagram in NetIQ Chariot

Unfortunately, CNet ProG2000L adapters showed no miracle. Their speed remained extremely low (although increased against Windows). And as was mentioned above, we couldn't test them with Jumbo frames enabled.

For some reason, the old dual-chip version of the Trendnet adapter (TRENDnet TEG-PCITX2) continues to show utterly bad results in this testing. Last time they were much better. Perhaps, it is all about being locally incompatible with the motherboard.

The 3Com 3C996B-T adapter on the bcm5700 driver has the best PI this time (on tg3, it has virtually the same speed but a twice-as-high CPU load). D-Link DGE-510T and ZyXEL Omni Lan PCI G1 showed good results too (but the latter adapter's driver is not stable enough). SysKonnet adadpters once again going neck-and-neck, occupy the place in the middle.

Peak performance results in NetPIPE, Linux

The NetPIPE utility tests the graph by constantly increasing the size of the packet for transmission. Thus, it enables us to know the bottlenecks where the adapter may have performance falls. As a result, we can estimate the adapter's peak speed (usually done on large-size packets) and speed changes in the range between the minimal and the maximal packet sizes.

adapters' peak speed with different Jumbo frame sizes, NetPIPE

CPU load with different Jumbo frame sizes during testing, NetPIPE

adapters' PI with different Jumbo frame sizes, the most typical diagram in NetPIPE

Most adapters (except CNet ProG2000L, ZyXEL Omni Lan PCI G1 and TRENDnet TEG-PCITX2) show a high peak speed.

The SysKonnect family has the highest CPU load with Jumbo frames disabled. Although the situation gets back to normal if we enable the frames, and the larger the size the better their results.

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.