Return to BSD News archive
Newsgroups: comp.os.386bsd.questions Path: sserve!newshost.anu.edu.au!munnari.oz.au!uunet!psgrain!percy!agora!implode!davidg From: davidg@implode.rain.com (David Greenman) Subject: Re: Ethernet [really TCP] performance measurement Organization: Delta Systems, Portland, OR. Message-ID: <C8L7A8.1v1@implode.rain.com> Date: Mon, 14 Jun 1993 01:10:55 GMT Lines: 152 Garrett Wollman writes: >You're probably comparing it with a 16-bit Ethernet adapter that has a >reasonable amount of memory. Me no got. The AT&T EN100 is an 8-bit >adapter (i.e., SLOW SLOW SLOW) with only 16k of memory. (I am certain >that I could handle 16k TCP windows if only I had 64k of memory...) The SMC 8013 is a 16 bit card, but it has just 16k of memory. As far as why you can't handle 16k windows - its because of a either a bug in the malloc code, or a bug in the driver. In the case of the malloc code, the problem is partially fixed in the patchkit. In the case of the driver, the original if_wd driver and the original if_ec driver had a bug that would crash the machine under heavy load. My new driver fixes these problems, and also deli- vers full ethernet performance with the 16bit WD/SMC boards. I'm not sure why Amancio has the 'low' performance of 'only' 985k/second, but considering how different our kernel sources are (Amancio uses NetBSD, I'm using "David's BSD"), this doesn't surprise me that much. Attached is a draft version of my driver's release notes. BTW, ttcp performance to localhost on my machine: ttcp-t: buflen=8192, nbuf=2048, align=16384/+0, port=5001 tcp -> localhost ttcp-t: socket ttcp-t: connect ttcp-t: 16777216 bytes in 5.40 real seconds = 3034.52 KB/sec +++ ttcp-t: 2048 I/O calls, msec/call = 2.70, calls/sec = 379.32 ttcp-t: 0.0user 3.3sys 0:05real 61% 0i+0d 0maxrss 0+0pf 2536+43csw The TCP window is set to 8k, 'implode' is a 486DX2/66. Now this is 20 times what you reported. 386DX/20's aren't _that_ slow. My little 386SX/25 gets about 600K/second. You really should have a look at your kernel source. -DG --- David Greenman davidg@implode.rain.com -------------------------------------------------------------------------- Release Notes for 'ed' Device Driver David Greenman, 24-May-1993 ------------------------------------ INTRODUCTION ------------ The 'ed' device driver is a new, high performance device driver supporting Western Digital/SMC 80x3 series (including 'EtherCard PLUS' 'Elite16') and the 3Com 3c503. All of the ethernet controllers use the DS8390 or 83C690 Network Interface Controller (NIC). The differences between the boards are in their memory capacity, bus width (8/16 bits), and special logic (asic) used to configure the shared memory and other things. Every effort has been made to conform to the manufactures' specifications for the NIC and asic. This includes both normal operation and error recovery. PERFORMANCE ----------- transmit -------- The 8390 doesn't provide a mechanism for chained write buffers, so it is very important for maximum performance to queue the next packet for transmission as soon as the current one has completed. On boards with 16k or more of memory, the shared memory is divided in a way that allows enough space for two full size packets to be buffered for transmission. When sufficient data is available for transmission, a packet is copied into the shared memory, the transmission is started, and then an additional packet is copied into the shared memory (to a different memory area). As soon as the first packet has completed, transmission of a second packet can then be started immediately - in less time than the 9.6uS interframe gap. This results in the highest performance possible from ethernet. Packets go out on the 'wire' with the following format: preamble dest-addr src-addr type data FCS intr-frame 64bits 48bits 48bits 16bits 1500bytes 32bits 96bits With 10Mbits/sec, each bit is 100nS in duration. All of the above fields, except for data are of fixed length. With full sized packets (1500 bytes), the maximum unidirectional data rate can be calculated as: 6.4uS + 4.8uS + 4.8uS + 1.6uS + 1200uS + 3.2uS + 9.6uS = 1230.4uS/packet = 812.74382 packets/second = 1219115.7 (1190k) bytes/second. With TCP, there is a 40 byte overhead for the IP and TCP headers, so there is only 1460 bytes of data per packet. This reduces the maximum data rate to 1186606 bytes/second. With TCP, there will also be periodic acknowledgments which will reduce this figure somewhat both because of the additional traffic in the reverse direction and because of the occasional collisions that this will cause. Despite this, the data rate has still been consistantly measured at 1125000 (~1100k) bytes/second through a TCP socket. In these tests, the TCP window was set to 16384 bytes. With UDP, there is less overhead for the headers, and with 1472 bytes of data per packet, a data rate of 1196358.9 (1168k) bytes/sec is possible. UDP performance hasn't been precisely measured with this device driver, but less precise tests show this to be true (measured at around 1135k/second). receive ------- The 8390 implements a shared memory ring-buffer to store incoming packets. The 8bit boards (3c503, and 8003) usually have only 8k bytes of shared memory. This is only enough room for about 4 full size (1500 byte) packets. This can sometimes be a problem, especially on the original WD8003E and 3c503. This is because these boards' shared memory access speed is also quite slow compared to newer boards - typically only about 1MB/second. The additional overhead of this slow memory access, and the fact that there is only room for 4 full-sized packets means that the ring-buffer will occassionally overflow. When this happens, the board must be reset to avoid a lockup problem in early revision 8390's. Resetting the board will cause all of the data in the ring-buffer to be lost - requiring it to be re-transmitted/received...slowing things even further. Because of these problems, maximum throughput on boards of this type is only about 400-600k per second. The 16bit boards (8013 series), however, have 16k of memory as well as much faster memory access speed. Typical memory access speed on these boards is about 4MB/second. These boards generally have no problems keeping up with full ethernet speed. The only problem I've seen with these boards is related to the (slow) performance of 386BSD's malloc code when additional mbufs must be added to the pool. This can sometimes increase the total time to remove a packet enough for a ring-buffer overflow to occur. This tends to be highly transient, and quite rare on fast machines. I've only seen this problem when doing tests with large amounts of UDP traffic without any acknowledgments (uni-directional). Again, this has been very rare. All of the above tests were done using a 486DX2/66, 486DX/33, 386DX/40, 8-9Mhz ISA bus, with Bruce Evans' high speed spl()/interrupt modifications, a high performance version of in_cksum.c from Bakul Shah, with tcp_sendspace/ tcp_recvspace set to 16k, and MCLBYTES set to 2048 (to allow full MTU sized packets). TCP tests were done with the 'ttcp' performance test utility, and also with FTP client/server. UDP tests were done with a modified version of ttcp (to work around a bug in 386BSD's UDP code related to queue depth), and also with NFS. KNOWN PROBLEMS -------------- 1) Early revision DS8390B chips have problems. They lock-up whenever the receive ring-buffer overflows. They occassionally switch the byte order of the length field in the packet ring header (several different causes of this related to an off-by-one byte alignment) - resulting in "shared memory corrupt - invalid length NNNNN" messages. The board is reset whenever these problems occur, but otherwise there is no problem with recovering from these conditions. 2) 16bit boards can conflict with 8bit BIOS or BIOS extensions (like the VGA). There is a work-around for this in the driver, however. The problem is that the ethernet board stays in 16bit mode, asserting its '16bit' signal on the ISA bus. This signal is shared by other devices/ROMs in the same 128K memory segment as the ethernet card - causing the CPU to read the 8bit ROMs as if they were 16bit wide. The work-around involves setting the host access to the shared memory to 16bits only when the memory is actually accessed, and setting it back to 8bit mode all other times. Without this work-around, the machine will hang whenever a reboot is attempted. 3) 16bit 3c503 boards seem to have a memory contention problem between the NIC and the on-board shared memory that causes periodic fifo-overruns. I'm looking for a work-around, but haven't found one yet. When this problem occurs, the 3c503 drops the packet. Other than re-tranmissions/reduced performance, no other effects of this problem have been seen.