Next Previous Contents

7. Technical Information

For those who want to understand a bit more about how the card works, or play with the present drivers, this information should be useful. If you do not fall into this category, then perhaps you will want to skip this section.

7.1 Programmed I/O vs. Shared Memory vs. DMA

If you can already send and receive back-to-back packets, you just can't put more bits over the wire. Every modern ethercard can receive back-to-back packets. The Linux DP8390 drivers (wd80x3, SMC-Ultra, 3c503, ne2000, etc) come pretty close to sending back-to-back packets (depending on the current interrupt latency) and the 3c509 and AT1500 hardware have no problem at all automatically sending back-to-back packets.

Programmed I/O (e.g. NE2000, 3c509)

Pro: Doesn't use any constrained system resources, just a few I/O registers, and has no 16M limit.

Con: Usually the slowest transfer rate, the CPU is waiting the whole time, and interleaved packet access is usually difficult to impossible.

Shared memory (e.g. WD80x3, SMC-Ultra, 3c503)

Pro: Simple, faster than programmed I/O, and allows random access to packets. Where possible, the linux drivers compute the checksum of incoming IP packets as they are copied off the card, resulting in a further reduction of CPU usage vs. an equivalent PIO card.

Con: Uses up memory space (a big one for DOS users, essentially a non-issue under Linux), and it still ties up the CPU.

Bus Master Direct Memory Access (e.g. LANCE, DEC 21040)

Pro: Frees up the CPU during the data transfer, can string together buffers, can require little or no CPU time lost on the ISA bus. Most of the bus-mastering linux drivers now use a `copybreak' scheme where large packets are put directly into a kernel networking buffer by the card, and small packets are copied by the CPU which primes the cache for subsequent processing.

Con: (Only applicable to ISA bus cards) Requires low-memory buffers and a DMA channel for cards. Any bus-master will have problems with other bus-masters that are bus-hogs, such as some primitive SCSI adaptors. A few badly-designed motherboard chipsets have problems with ISA bus-masters.

7.2 Performance Implications of Bus Width

The ISA bus can do 5.3MB/sec (42Mb/sec), which sounds like more than enough for 10Mbps ethernet. In the case of the 100Mbps cards, you clearly need a faster bus to take advantage of the network bandwidth. a 33MHz 32 bit PCI bus can do 133MB/sec which isn't enough for GigE.

ISA Eight bit and ISA 16 bit Cards

You probably will have a hard time buying a new ISA ethercard anymore, but you can probably still find some surplus or obsolete cards suitable for ``home-ethernet'' systems. If you want to really go retro, you can even use an old half slot 8 bit ISA card, but note most of them are 10Base-2.

Some 8 bit cards that will provide adequate performance for light to average use are the wd8003, the 3c503 and the ne1000. The 3c501 provides poor performance, and these poor 15 year old relics of the XT days should be avoided. (Send them to Alan, he collects them...)

The 8 bit data path doesn't hurt performance that much, as you can still expect to get about 500 to 800kB/s ftp download speed to an 8 bit wd8003 card (on a fast ISA bus) from a fast host. And if most of your net-traffic is going to remote sites, then the bottleneck in the path will be elsewhere, and the only speed difference you will notice is during net activity on your local subnet.

32 Bit PCI (VLB/EISA) Ethernet Cards

Obviously a 32 bit interface to the computer is a must for 100Mbps and higher networks. If you get into GigE, then the 133 megabyte/sec PCI bus (for 33MHz 32 bit PCI) will still be your limiting factor.

But an older 10Mbs network doesn't really require a 32 bit interface. See Programmed I/O vs. ... as to why having a 10Mbps ethercard on an 8MHz ISA bus is really not a bottleneck. Even though having a slow ethercard on a fast bus won't necessarily mean faster transfers, it will usually mean reduced CPU overhead, which is good for multi-user systems.

7.3 Performance Implications of Zero Copy

As network data is sent or received, you can easily imagine it being copied to/from the application into kernel memory and from there being copied to/from the card memory. All this data movement takes time and CPU resources. As hinted above in the Bus Master DMA section, a properly designed card can cut down on all this copying, and the most ideal case would be zero copy of course. With some of the modern PCI cards, zero copy is possible by simply pointing the card at the data and essentially saying "get it yourself." If maximum performance with minimum server load is important to you then check to see if your hardware and driver will support zero copy.

7.4 Performance Implications of Hardware Checksums

There is no guarantee that your data will travel from computer A to computer B without being corrupted. To make sure the data is OK, the sender adds up all the numbers that make up your data, and sends this checksum along as well. The receiver recomputes this checksum and compares it to the one the sender computed. If the two don't match, the receiver knows that the data has been corrupted and it will reject the bad data.

Computing these sums takes time and extra load on the main computer. Some of the more fancy cards have the ability to do these Rx and/or Tx sums in hardware, which allows the main CPU to offload this task to the card.

Cards that require a data copy don't benefit as much from hardware checksums, since the sum operation can be combined into the copy for only a minimal additional overhead. Hence hardware Tx checksums are only used in zero copy (i.e. applications using sendfile()) situations, and so hardware Rx checksums are currently more useful.

Note that a reasonable computer can saturate a 100BaseT link even when doing the copy and checksum itself, so zerocopy/hw-checksum will only show up as decreased CPU use. You would have to go to GigE to see a speed increase.

7.5 Performance Implications of NAPI (Rx interrupt mitigation)

When a card receives a packet from the network, what usually happens is that the card asks the CPU for attention by raising an interrupt. Then the CPU determines who caused the interrupt, and runs the card's driver interrupt handler which will in turn read the card's interrupt status to determine what the card wanted, and then in this case, run the receive portion of the card's driver, and finally exits.

Now imagine you are getting lots of Rx data, say 10 thousand packets per second all the time on some server. You can imagine that the above IRQ run-around into and out of the Rx portion of the driver adds up to a lot of overhead. A lot of CPU time could be saved by essentially turning off the Rx interrupt and just hanging around in the Rx portion of the driver, since it knows there is pretty much a steady flow of Rx work to do. This is the basic idea of NAPI.

As of 2.6 kernels, some drivers have a config option to enable NAPI. There is also some documentation in the Documentation/networking directory that comes with the kernel.


Next Previous Contents