11/70 Loads

[utzoo!decvax!ucbvax!unix-wizards]

>From mike.bmd70 at BRL Tue Sep 15 19:21:50 1981
Lt. Breitling -
	We are running an enhanced V6 UNIX on 2 11/70s here, and have
some of the following performance observations:

1)  The NCP code does not seem to use much of the CPU, cycle wise.  The
    user programs are not really much worse.  You do pay a penalty in
    core usage, because the network software does occupy a fair amount
    of core...  3 TELNETs, the NCP, ServerTelnetDaemon, FTPServerDaemon,
    and an active FTP or two takes almost 600 Kb.  (This observation
    made with memory lights & memory utilization analyzer).  This does not
    grow linearly with load, however, thanks to the "pure text" mechanism.

2)  Our primary 11/70 delivers sub-second (0.1 -> 0.2 s) response time for
    most requests (editor commands, WHO, MOUNT, etc) which do not demand
    significant I/O from the moving head disks.  We can deliver this with
    30 users on the system (1/3 doing program development, 2/3 editing and
    TROFF).  However, we have a rather special hardware configuration which
    aids this:
		ROOT		4096 blocks of BULK CORE (RF-11 interface)
		SWAP		4872 blocks of RK-11
		/tmp		4096 blocks of BULK MOS
		 *		3 CDC 80 Mb drives, Massbus interface
	      -MEMORY-		1.5 Mb

3)  "Number cruncher" processes tend to get little or no CPU time durring
    prime time;  they just sit around and wait for spare cycles.  This
    is due to our "High Performance Scheduler".  Gives great interactive
    response, and prevents "batch" work from interfering with interactive
    terminals.

4)  Although 1.5 Mb of main memory does seem to be sufficient now, when
    we had only 0.5 Mb, we did observe swapping rates of up to 10/sec
    before performance began to suffer too much.  Swapping is to be avoided.

5)  With the addition of another MASSBUS disk interface and more drives,
    we could probably support 48-55 users with the same performance,
    although the CPU would probably start to run out of steam at this point.

6)  I could go on and on about performance issues.  If there is interest,
    we could put together a somewhat longer letter....  What interests
    people?

7)  We plan to put UNIX on the HEP (Heterogeneous Element Processor),
    a multi-CPU machine, where the raw speed of each CPU element is 10 MIPS
    and only costs $250K - $500K (not sure).  Our initial configuration
    will be 40 MIPS.  (From the people who brought you ENIAC, ...HEP!)
    We think it will make a superb interactive facility.

				Cheers,
				 -Mike