fedorkow@BBN-VAX.ARPA (Guy Fedorkow) (10/01/85)
In the process of evaluating computing environments for running Magic and other Berkeley VLSI tools at BBN, I have compared the performance of our VAX 11/785 and Sun Microsystems workstations. This memo describes the method used and the results obtained, and as such, it may be of interest to others contemplating the purchase of hardware to support VLSI work. Questions or comments may be directed to fedorkow@bbn-vax A visit to Sun Microsystems in Lexington MA Guy Fedorkow, BBN, Sept 23, 1985 After several attempts, I finally had a chance today to use a sun-3/160 workstation to run a Magic benchmark. It was worth the wait; the new Sun-3 systems seem to be real performers. The system that I used was a Sun 3/160C, a color workstation with 140 Mb of local disk, 4Mb of main memory, and a 68020/68881 CPU. The machine had the sun graphics processor, but it apparently has very little impact on the current version of Magic. The tests that I ran involved a series of manipulations of the BBN Butterfly switch chip, excluding pads, etc. This is an nmos chip, using just under 2,000 transistors. The basic intent of the test was to isolate the performance contributions from graphics display vs compute activity, for small and large chips. The 2,000 transistor switch chip was flattened to simulate the effects of a large chip. Results in this test sequence are compared to a VAX 11/785. Tests were run on the VAX early in the morning, at a time when serious users were still asleep. No other precautions were taken to ensure uniform vax loading. Partial results are also presented for the same test run on a sun 2/160, the machine without the new 68020 board. This set of tests could not be completed due to changes in a new version of magic. Timing measurements are expressed as minutes:seconds. Don't try interpret values to five decimal places -- all measurements were done with an ordinary stop watch. The Tests Command Vax Sun 3/160 Sun 2/160 /* cause the entire chip to be displayed */ :expand all 3:54 1:13 3:40 /* check the entire chip - a test of compute speed */ :drc check 0:42 0:31 1:28 /* flatten the hierarchy */ :yank 0:30 0:28 3:45 :load new /* stuff into a new cell */ :stuff 0:40 1:25 it broke /* note the effects of paging are beginning to show */ /* write a 330 Kbyte file */ :cif write 1:00 2:15 :extract all 0:36 3:00 Summary The sun 68020 implementation seems to be fast indeed. Pictures are drawn about 3 times faster than the VAX with an AED-512. Computing seems to on a par with the 785, as long as the core image is small enough to fit into ram. Disk I/O is still slow -- once paging sets in, the sun slows down a lot. The same result has been observed in previous tests with a 2 Mbyte sun without the 68020 upgrade. In that case, performance with magic fell by a factor of ten once serious paging started to happen.