Here are some tips that you can use if you are suffering from low ethernet throughput, or to gain a bit more speed on those ftp transfers.
ttcp.c program is a good test for measuring
raw throughput speed. Another common trick is to do
ftp> get large_file /dev/null where
large_file is > 1MB and residing in the buffer
cache on the Tx'ing machine. (Do the `get' at least
twice, as the first time will be priming the buffer
cache on the Tx'ing machine.) You want the file in
the buffer cache because you are not interested in
combining the file access speed from the disk into
your measurement. Which is also why you send the
incoming data to
/dev/null instead of onto
Even an 8 bit card is able to receive back-to-back packets without any problems. The difficulty arises when the computer doesn't get the Rx'd packets off the card quick enough to make room for more incoming packets. If the computer does not quickly clear the card's memory of the packets already received, the card will have no place to put the new packet.
In this case the card either drops the new packet, or writes over top of a previously received packet. Either one seriously interrupts the smooth flow of traffic by causing/requesting re-transmissions and can seriously degrade performance by up to a factor of 5!
Cards with more onboard memory are able to ``buffer'' more packets, and thus can handle larger bursts of back-to-back packets without dropping packets. This in turn means that the card does not require as low a latency from the host computer with respect to pulling the packets out of the buffer to avoid dropping packets.
Most 8 bit cards have an 8kB buffer, and most 16 bit cards have a 16kB buffer. Most Linux drivers will reserve 3kB of that buffer (for two Tx buffers), leaving only 5kB of receive space for an 8 bit card. This is room enough for only three full sized (1500 bytes) ethernet packets.
As mentioned above, if the packets are removed from the card fast enough, then a drop/overrun condition won't occur even when the amount of Rx packet buffer memory is small. The factor that sets the rate at which packets are removed from the card to the computer's memory is the speed of the data path that joins the two -- that being the ISA bus speed. (If the CPU is a dog-slow 386sx-16, then this will also play a role.)
The recommended ISA bus clock is about 8MHz, but many motherboards and peripheral devices can be run at higher frequencies. The clock frequency for the ISA bus can usually be set in the CMOS setup, by selecting a divisor of the mainboard/CPU clock frequency. Some ISA and PCI/ISA mainboards may not have this option, and so you are stuck with the factory default.
For example, here are some receive speeds as measured by the TTCP program on a 40MHz 486, with an 8 bit WD8003EP card, for different ISA bus speeds.
ISA Bus Speed (MHz) Rx TTCP (kB/s) ------------------- -------------- 6.7 740 13.4 970 20.0 1030 26.7 1075
You would be hard pressed to do better than 1075kB/s with any 10Mb/s ethernet card, using TCP/IP. However, don't expect every system to work at fast ISA bus speeds. Most systems will not function properly at speeds above 13MHz. (Also, some PCI systems have the ISA bus speed fixed at 8MHz, so that the end user does not have the option of increasing it.)
In addition to faster transfer speeds, one will usually also benefit from a reduction in CPU usage due to the shorter duration memory and I/O cycles. (Note that hard disks and video cards located on the ISA bus will also usually experience a performance increase from an increased ISA bus speed.)
Be sure to back up your data prior to experimenting with ISA bus speeds in excess of 8MHz, and thouroughly test that all ISA peripherals are operating properly after making any speed increases.
Once again, cards with small amounts of onboard RAM and relatively slow data paths between the card and the computer's memory run into trouble. The default TCP Rx window setting is 32kB, which means that a fast computer on the same subnet as you can dump 32k of data on you without stopping to see if you received any of it okay.
Recent versions of the
route command have the ability
to set the size of this window on the fly. Usually it is only
for the local net that this window must be reduced, as computers
that are behind a couple of routers or gateways are `buffered'
enough to not pose a problem. An example usage would be:
route add <whatever> ... window <win_size>
win_size is the size of the window you wish to
use (in bytes). An 8 bit 3c503 card on an ISA bus operating
at a speed of 8MHz or less would work well with a window
size of about 4kB. Too large a window will cause overruns
and dropped packets, and a drastic reduction in ethernet
throughput. You can check the operating status by doing
cat /proc/net/dev which will display any
dropped or overrun conditions that occurred.
Some people have found that using 8 bit cards in NFS clients causes poorer than expected performance, when using 8kB (native Sun) NFS packet size.
The possible reason for this could be due to the difference in on board buffer size between the 8 bit and the 16 bit cards. The maximum ethernet packet size is about 1500 bytes. Now that 8kB NFS packet will arrive as about 6 back to back maximum size ethernet packets. Both the 8 and 16 bit cards have no problem Rx'ing back to back packets. The problem arises when the machine doesn't remove the packets from the cards buffer in time, and the buffer overflows. The fact that 8 bit cards take an extra ISA bus cycle per transfer doesn't help either. What you can do if you have an 8 bit card is either set the NFS transfer size to 2kB (or even 1kB), or try increasing the ISA bus speed in order to get the card's buffer cleared out faster. I have found that an old WD8003E card at 8MHz (with no other system load) can keep up with a large receive at 2kB NFS size, but not at 4kB, where performance was degraded by a factor of three.
On the other hand, if the default mount option is to use 1kB size and you have at least a 16 bit ISA card, you may find a significant increase in going to 4kB (or even 8kB).