Re: Network file transfer rate comparisons

["David A. Moon" <Moon@SCRC-STONY-BROOK.ARPA>]

Just to set the record straight, based on my knowledge of the underlying
disk hardware and software:

    Date: Thu, 26 Jan 89 10:09 EST
    From: Len Moskowitz <Len@HEART-OF-GOLD>
 
    It seems to me that your performance test will reflect disk access
    times just as much as network performance.  I tried it with a 3675
    accessing its own two large drives (a 515 and a 470) using NFILE and
    found the following performance:

Unless you did something magic to get the 3675 to use NFILE to talk to
itself, instead of going direct, you weren't using NFILE and weren't
using the network.  Naturally going direct is faster.  You really need
to use two machines to perform these tests.

    Date: Thu, 26 Jan 89 15:29 EST
    From: miller@CS.ROCHESTER.EDU (Brad Miller)
    
    Note that 3670s have a much faster SMD access time than the 50. Part of the
    reason is software, part of the reason is that the channel is 16 bits wide
    on the L machines, and only 8 bits wide on the N machines.
    
It's true that the disk driver software on the 3670 is better than on the
3650.  The hardware is also different, whether it's better is hard to say.
The rest of this (the 8-bit/16-bit stuff) is not true.  The only referent
for those numbers I can find is the bus width of the FEP microprocessor.

Since raw disk speeds are in the 1 to 3 million bytes per second range, and
the numbers quoted here are in the 20 to 80 thousand bytes per second range,
I doubt that raw disk speed has anything to do with the measured performance
of any of these file systems and network file access paths.  A few days
with the Metering Interface (the one introduced in 7.2) would probably
reveal some of the roots of the measured performance.  Raw disk speed does
affect paging performance pretty noticeably, though.

    So, 367x's make faster file servers than 3650s (if you can afford the
    maintainance).

I haven't tried to compare the speed of those two machines as file servers,
but I suspect the story is much more complicated than anything that's been
presented so far.  There are a lot of variables.  I do know that as a file
server shared between more than a couple of people, LMFS is generally cpu
bound.  It wouldn't surprise me if the 3675 comes out faster, but it
might not be for the reason you think.

I want an XL400 file server.