jqj@cornell.UUCP (05/21/85)
From: jqj (J Q Johnson) To reiterate a request that has already appeared on Unix Wizards, can anyone out there provide information on MicroVAX II performance with multiple users under Unix? Granted, the MV II is a fast, nifty machine for a small number (1 to 3?) of users; can it replace a 780 (750?) for ordinary timesharing? It certainly won't replace my 780 with 2GB of disk and 40 to 50 users (at least not this year), but how would it perform in a configuration with, say, 5MB memory, 5 RD53 disks, and 12 simultaneous timesharing users? Some issues: The RD53 is a fairly slow disk (RD52 is too slow to even consider in this context). The controller for A-series disks won't be available till late this year (or, more likely, early next year). Will lots of spindles make up for slower average access? Let's ignore the issue of whether 350MB is enough disk and simply worry about performance for a while. The MV II uses the Q bus for all I/O. Is this a potential bottleneck, particularly in an environment with moderately high Ethernet traffic competing for the bus? My experience with 780s seems to be that they are very happy to have their disks sitting on Massbus or SI controllers (or on a CI). If I were configuring a 780 with UDA-50, I think I'd want to hang it on a dedicated Unibus. Can I install a second Q bus on the MV II? How smart is the controller for the RD53? In a typical disk transfer, how much work does the CPU have to do compared, for example, to a UDA50? How wide is the DMA path from Qbus to memory? Unless memory deceives me, the VAX/750 has only 4 registers for controlling byte DMA transfers, which makes them a very scarce resource, especially if for speed your drivers want to dedicate one or two to each DMA device; 780 has 64, which is plenty. How many has the MV II? How big is the cache, and how fast is main memory? Floating point performance: I gather that the MV II supports G and H format fp numbers. How about D format? In hardware or in emulation? How fast? Note that binaries from existing C programs will do all their calculations in D format, so if D format is slow one would at least want to recompile using a compiler that generates G format; what does the distributed Ultrix C compiler use? What other potential bottlenecks exist in the MV II architecture?
dlm@piggy.UUCP (Daryl Monge) (05/23/85)
I have been very impressed by the flood of incorrect information in this group concerning the uVAX-II. For example ( from DEC publications): The RA81 average access time is 36ms. The RM05 average access time is 38.3ms. the RD53 average access time is 38.3ms. This is slow? Also, the performance benchmarks we have heard about from prerelease users have involved real world applications. I consider this to be much more useful information than the amount of cache, etc. I don't even care about the cache if my application performs nearly equal to a 780. Lastly, several persons have already written in this newsgroup that the QBUS is 3.3 MBytes/sec (3 times faster than a 780 UNIBUS). The CPU uses 6.6MBytes/sec on a special memory bus and the memory is a 10MByte/sec system. 6.6+3.3<10. If your UNIX or VMS system can generate 3.3MBytes/sec on data traffic on the bus, I sure would like to hear about it! Daryl Monge AT&T Bell Labs Holmdel, NJ ..!ihnp4!piggy!dlm
dan@rna.UUCP (Dan Ts'o) (05/26/85)
In article <> dlm@piggy.UUCP (Daryl Monge) writes: >I have been very impressed by the flood of incorrect information >in this group concerning the uVAX-II. > >For example ( from DEC publications): > >The RA81 average access time is 36ms. >The RM05 average access time is 38.3ms. >the RD53 average access time is 38.3ms. > >This is slow? > >Also, the performance benchmarks we have heard about from prerelease users >have involved real world applications. I consider this to be much more >useful information than the amount of cache, etc. I don't even care about >the cache if my application performs nearly equal to a 780. > >Lastly, several persons have already written in this newsgroup that the >QBUS is 3.3 MBytes/sec (3 times faster than a 780 UNIBUS). The CPU uses >6.6MBytes/sec on a special memory bus and the memory is a 10MByte/sec >system. 6.6+3.3<10. >If your UNIX or VMS system can generate 3.3MBytes/sec on data traffic on >the bus, I sure would like to hear about it! Well I don't want to add more misinformation. The RD53 may be a perfectly fine disk for some applications. But two more considerations may make the RD53 "slow". 1) Transfer rate. The RD52, like the RD5[012], is rated at 5Mbit/s which is about 600kbytes/sec. The RM03/5 are rated at 1.2Mbytes/sec. The Fuji Eagle is rated at 1.8Mbytes/sec. Thus the RD53's transfer rate is similar to the RL02 and the RK07 and only half that of the RM's and a third of the Eagle's, our standard VAX disk. 2) Controller performance. The RD53 uses the RQDX2, an upgraded version of the RQDX1 (RD5[012]). I have an RQDX1 and it is incredibly slow. It uses a T-11 chip (of FALCON 11/21) which is similar in performance to a Z80. There are status LED's on the board. You can watch fractions to seconds go by as the controller goes about its work. To be fair, the RQDX2 is supposed to be somewhat better. But it is the same technology and no match for the bitslice SMD controllers on the MASSBUS. Actually I have a Dilog controller on a Qbus driving an Eagle. It has to interleave (2:1 ?) to slow down the eagle but its still quite a performer. I would serious consider getting an Eagle-like drive and a decent controller for the uvax II and use the RD5[0-3] for booting and DEC-appeasement.
hedrick@topaz.ARPA (Chuck Hedrick) (06/02/85)
I just came back from DECUS. One interesting talk was about implementing Ultrix on new processors. It reported the results of the first 2 of 4 such projects that they are working on this year (the 8600 and the microVAX II -- when asked what the other two were, they just smiled). In the Q & A, I asked the question "What is likely to be the difference in performance between a microVAX II and a 780?" Here is roughly what he said. (This is from memory, by the way, and I don't claim to have an eidetic memory. I am integrating the answers to a couple of questions.) We get asked this a lot. The first thing we always try to make clear is A MICROVAXII IS NOT A 780. It is the greatest thing you would ever want to see as a single-user system. I've got one under my desk, and the day somebody tries to take it, they'll do it over my dead body. But we have never claimed that it will do everything a 780 will do. In fact in some cases, it isn't even a 750. It is not a matter of CPU speed. The chip is a real screamer. But there are two critical differences between a microVAX and a 780: 1) The microVAX II is a single-bus system. We normally configure 780's with more than one Unibus adapter. We try to put bus hogs such as the UDA50 and DEUNA on separate buses. 2) The combination of controller and disks simply do not have as much speed as a UDA50 with an RA81. We have played with a hypothetical controller similar to the UDA50, and it seems to get about twice the throughput of even a system with 2 RD53's. [end of detailed summary] The responder seemed to think that it would be unlikely that a microVAX, as currently configured, would be able to support 16 users. He implied that with the KDA50 (the hypothetical Q-bus version of the UDA50), it would. Obviously it is unfair to expect a single number for user capacity, since it depends so much upon what is going on. But it sounded to me like 8 might be feasible.