tcs%usna.uucp@BRL-BMD.ARPA (02/05/85)
Mike Muuss (mike@brl) and I (tcs@usna) briefly tested the Excelan Unibus Ethernet controller (Exos 204) in link controller mode on two Vax 11/780 processors. Below is the performance matrix we obtained from the Excelans, a pair of Interlans (NI1010), and a pair of PCL-11s. The speed of the PCL was approximately 1M bit/sec between the two systems. The test program was ttcp, modified to run on 4.2. Ttcp is a program that can transmit or receive data over network connections. In transmitter mode, it can send a user specified number of 1K byte zero filled packets. In receiver mode, the packets are dumped one by one into a single buffer, overwriting the previous packet. This is useful for benchmarking network performance between two processes. Different modes allow the transmitter and receiver to read and write data on stdin and stdout. This routine is available to anyone interested. The writing cpu was BRL's dual vax (VGR) which we ran in both dual and single processor configurations. The receiving cpu was a standard 780 (TGR). Both systems have the new MS780E memory controllers with at least 6Mby of memory. VGR was running multi-user but we were the only users on the system. TGR was running single user. We tested the links by sending 1Mby of data as 1K byte packets. We tested with trailers on and off. We found it interesting that the system load reduction from using trailers is a lot less than expected. The pcl doesn't have a local network packet so there are no headers or trailers. We did run across an anomoly where the system times for the pcl dropped to 5-8%. It did this three consecutive times (two documented below). We ran some other network & disk routines (fsck, rlogin, ftp) and the system times jumped back up. Note that the usage was similar on both machines. This would seem to rule out any alignment with the system clock which might hide the utilization from process accounting. Anyone have any ideas? The excelan was setup to burst 4 words per unibus dma. It is jumper configurable to 2, 4, 8, or 16 word bursts. Excelan's additional CPU overhead per kilobyte is due in part to handling the message queues used to communicate with the board (similar to the DEUNA). We did not examine the driver for possible improvements in performance. The cpu configurations of VGR are noted as 'dual' and 'single'. VGR (Transmitter) Interface Usr Sys % cpu Real Kby/S Kby/Cpu-Sec --------------------------------------------------------------------- dual,ex0,-trailers .5 6.5 40 17 60.2 150 dual,ex0,trailers .5 6.6 42 16 64 152 dual,il0,-trailers .3 5.4 25 22 46.5 186 dual,il0,trailers .4 5.6 24 24 42.6 178 dual,pcl0 .5 4.9 24 18 56.8 237 single,pcl0 .2 3.9 22 19 53.8 244 single,pcl0 .0 1.7 8 19 53.8 673 single,pcl0 .0 1.5 8 19 53.8 673 >>>>> ran fsck, rlogin, ftp <<<<<< single,pcl0 .1 5.4 28 20 51.2 182 single,pcl0 .6 5.4 32 18 56.8 178 single,il0,trailers .3 5.8 24 25 40.9 170 single,ex0,trailers .5 6.0 35 18 56.8 162 single,ex0,trailers .4 6.2 36 18 56 156 single,ex0,-trailers .6 6.0 38 17 60.2 158 single,ex0,-trailers .6 5.9 38 17 60.2 158 TGR (Receiver) Interface Usr Sys % cpu Real Kby/S Kby/Cpu-Sec --------------------------------------------------------------------- dual,ex0,-trailers .1 8.7 45 17 60.2 133 dual,ex0,trailers .0 10.3 59 16 64 108 dual,il0,-trailers .1 5.8 25 22 46.5 186 dual,il0,trailers .1 5.7 23 24 42.6 185 dual,pcl0 .0 5.3 24 18 56.8 237 single,pcl0 .0 4.6 24 19 53.8 244 single,pcl0 .0 1.1 5 19 53.8 1076 single,pcl0 .0 1.3 6 19 53.8 896 >>>>> ran fsck, rlogin, ftp <<<<<< single,pcl0 .0 6.6 31 20 51.2 165 single,pcl0 .0 6.7 35 18 56.8 162 single,il0,trailers .1 5.4 21 25 40.9 194 single,ex0,trailers .1 10.1 54 18 56.8 105 single,ex0,trailers .0 9.8 54 18 56 104 single,ex0,-trailers .1 9.3 53 17 60.2 114 single,ex0,-trailers .1 9.2 53 17 60.2 114 -tcs Terry Slattery U.S. Naval Academy 301-267-4413 ARPA: tcs@brl-bmd UUCP: decvax!brl-bmd!usna!tcs
jbn@wdl1.UUCP (02/08/85)
Note that this is all for the Excelan card in dumb Ethernet controller mode, not with its onboard IP/TCP. We have both Excelan and CMC boards in a VME cage, as well as an Excelan board in a Multibus cage on a Masscomp, and are waiting for software delivery on both. CMC claims that 1Mb/sec is possible in end-to-end FTP operations using their card; we look forward to trying this.