cak@Purdue@sri-unix (09/01/82)
From: Chris Kent <cak at Purdue> Date: 26 Aug 1982 14:04:58-EST Berkeley has one working for 4.1x. Rob Gurwitz is working on one for his TCP/IP. George Goble has a 'raw' one that just spits onto and reads off of the bus (contact mosher@Berkeley, gurwitz@bbn-unix and ghg@purdue, respectively). George Gobel reports being able to send a packet to the Interlan board only every 8 milliseconds, mainly due to the z80 on the board being to slow. The newer release of the board (due soon) should have a 30% speedup, according to Interlan. Still not great. chris
croft@SRI-TSC@sri-unix (09/04/82)
From: Bill Croft <croft@SRI-TSC> Date: 1 Sep 1982 at 1648-PDT Here is a note describing tests recently run by INTERLAN to determine the transmit performance of their NI1010 Unibus Ethernet Controller. The basic result is about 2 to 3 megabits per second maximum transmit rate. As far as I know, the 3COM board is in the same range. Of course what really counts is end to end thruput and this is more dependent on host buffering strategies and protocol efficiency. We'll have to wait for Berkeley's 4.2BSD network device benchmarks for an accurate comparison. The Interlan board does have the advantage of DMA operation, so it should be less load on the host CPU. (The NI1010A figures shown below are for the new bit-slice version of the board; the original NI1010 contains a Z80) --- TRANSMIT PERFORMANCE OF THE NI1010 ON A STANDALONE PDP-11/24 On an 11/24, the NI1010 was driven by a small MACRO-11 program that repetitively DMAed an Ethernet packet from UNIBUS memory to the controller which then transmitted it onto an idle Ethernet. The following tests consisted of measuring the time required to send 10000 packets of the given fixed size: Transmit Packet Size 64 bytes 200 bytes 800 bytes 1518 bytes -------- --------- --------- ---------- Packets 1010 699 290 171 per sec NI1010 .52 Mbps 1.12 Mbps 1.86 Mbps 2.07 Mbps transmit thruput NI1010A 1 Mbps 2.7 Mbps transmit thruput
cak@Purdue@sri-unix (09/08/82)
From: Chris Kent <cak at Purdue> Date: 5 Sep 1982 21:08:30-EST Mike O'Dell (mo@lbl-unix) would you please publish the numbers you've been getting with your proNET setup? The world is waiting. chris
mo@LBL-UNIX@sri-unix (09/10/82)
From: mo at LBL-UNIX (Mike O'Dell [system]) Date: 8 Sep 1982 09:35:08-PDT Ahem! The tests we have done involved running 4.1a to 4.1a between 780's over the pronet ring. We are seeing about 100 kbytes/sec user-process-to-user-process before the Vaxen burn up. There is no evidence of any wire-limiting going on. IP/TCP is a pig and there ain't no way around it, and the 4.1a TCP is probably the fastest around. As soon as Sam gets Delta-T working, I will try it and see what happens. Should be able to go a lot faster. In the near future, we will be connecting a logic analyzer to the hardware and be taking some detailed "this many microseconds to do that" kind of measurements. We do know that because the interface is only single-buffered on the input side, it can't do back-to-back packets, but that is a small loss compared to the braindamage of existing Ethernet controllers. Hope this information, what there is of it, is useful. -Mike
lwa%MIT-CSR@MIT-Multics@sri-unix (09/19/82)
Date: 15 Sep 1982 1652-EDT (Wednesday) Well, we have about 15 VAX-11/750's on a Pronet ring here, running Gurwitz' 4.1 TCP/IP code, and we are seeing file transfers (in image mode) around 250-300 KBits/second. I'm not sure I understand why you're seeing such poor performance; do you have the latest distribution from Rob? I understand there were some improvements in the "sys.8" code... BTW, the IP implementation itself runs somewhat faster. Using a locally- designed protocol for updating remote bitmapped displays (running directly on top of IP), we see data transfer rates of better than 400KBits/second. This is still not superb, but it's not too unsatisfactory. -Larry -------