NetNews Usenet Archive 1992 #19/NN_1992_19.iso/spool/comp/parallel/2033

Text File | 1992-09-02 | 8.8 KB | 240 lines

Newsgroups: comp.parallel
Path: sparky!uunet!gatech!hubcap!fpst
From: [e-mail] (H.V. Sreekantaswamy)
Subject: SUMMARY: Communication startup costs
Message-ID: <[e-mail]>
Apparently-To: [e-mail]
Followup-To: comp.parallel
Sender: [e-mail] (Usenet News)
Organization: Computer Science, University of B.C., Vancouver, B.C., Canada
Date: Wed, 2 Sep 92 23:27:37 GMT
Approved: [e-mail]
Lines: 226

Here are the replies I got for my request about communication startup
costs on various distributed memory machines. I have only included
the replies that had any information or pointers.

Thanks to everyone who replied.
--Swamy

------------------------------------------------------------------------
H.V. Sreekantaswamy            E-mail:    [e-mail]
Dept. of Computer Science        Tel:    [phone]
University of British Columbia            [phone] (Home)
Vancouver, B.C., Canada V6T 1Z2
------------------------------------------------------------------------

Original Request:
----------------
I am interested in knowing typical values for communication startup
costs on various distributed memory machines such as CM-5, nCUBE,
Intel IPSC, Delta, Paragon, etc. What I mean by communication startup
cost is the time it takes to begin the actual message transmission
from the time the user program makes the appropriate function call.
This should include all the operating system or runtime system overheads
in initiating the communication in addition to hardware setup time.
I would appreciate if anyone can point to any papers/reports that
discuss about these overheads or give the values for any system.

If there is sufficient interest, I will post a summary of replies.

Thanks,
--Swamy
*******************************************

From:          <[e-mail]>


See Gordon Bell's article in the August CACM.  He provides some of the
communication cost you seem to want.

-------------------

From:          <[e-mail]>

You could look at "low-latency message communication support
for the ap1000" in this year's ACM computer architecture conference.

----------------------
From:          Roland Ruehl <[e-mail]>
>Organization: IIS, Swiss Fed. Inst. of Technology


Suppose the delay T_del of a communication from cell to cell of a message
of size M is decomposed into latency and bandwidth contributions:

    T_del = T_lat + M / B

and T_c is the time to compute locally a DAXPY iteration.

The following table is an extract from my thesis (all times are in
musec and B in MBytes/sec):

Machine        |      T_c    |       T_lat    |    B    | Reference
----------------+---------------+---------------+---------------+-----------
  nCube/2    |      2.56    |     154    |     1.7    | (1 + 2)
  iPSC/i860    |      0.31    | M<=100:  75    |     2.5    | (1 + 2)
        |        | M >100: 136    |        |
  Delta        |      0.31    |     12.5    |     2.1    | (1 + 2)
  CM5 (scalar)    |      0.5    |     >= 3    |   5 ... 20    |   (3)
      (vector)    |      0.18    |     >= 3      |   5 ... 20    |
  AP1000    |      1.00    |     20    |     8.5    |   (4)

References:

1) LINPACK performance measured by Dongarra
2) Tutorial by Dongarra at Supercomputing 91
3) CM5 Performance profile
4) my thesis

I would be interested in getting the numbers for other new machines
(like the Paragon or Parsytec GC) if you can get them.

---------------------------------------------------------------------------

Roland Ruehl                                     Phone:     [phone]
Integrated Systems Laboratory                    Telex:     53 178 ethbi ch
ETH-Zuerich                                    Telefax:     [phone]
Switzerland                                     E-mail:   [e-mail]

-------------------------

From:          <[e-mail]>

An example you haven't mentioned are transputers. The current generation
takes ~20 cycles for the appropriate instruction, add a few cycles for
loading operands - say 25 cycles at 25 MHz = one microsecond. Data transfer
is done by a dedicated DMA machine, which also reschedules the process at
the end of a transfer. As transputers use synchronous communication, a
message can be any length. The new generation, due out at the end of the year,
will have automatic demultiplexing/multiplexing for the physical links and
a worm-hole routing chip; IN now takes 16 cycles - lets say 20 cycles at
50 MHz = 400 nanoseconds.

-- Jan Vorbrueggen, [e-mail]
------------------------
From:          <[e-mail]>
>Organization: Dartmouth College, Hanover, NH


 A paper by Marco Annaratone which I think came out sometime this year
 in IEEE TPDS has some latency numbers. The machine they report about
 are Ncube, Intel ipsc/860 and the Paragon.

 If you don't get hold of it, send me e-mail after 3 weeks. I am in the
 process of moving right now.


 --Deb Banerjee

--------------------------------------------
From:          <[e-mail]>

Swamy

Your definition of communications startup time is poor. It allows someone
to quote a figure which is just the time taken to put data into a FIFO. It
doesn't require any processing at the destination. The only reason for
communicating is to have a second process cooperate with the first so you
should be measuring the time taken to have something occur on a
remote processor.

The standard way of doing this is the so called "ping-pong" benchmark in which
process 1 sends a message to process 2 who replies. The message latency is half
the roundtrip time as measured on 1.


--

Best Wishes
Duncan Roweth

-------------------------------------------------------------
|  Meiko Limited                Phone : [phone]      |
|  650 Aztec West               FAX   : [phone]      |
|  Bristol BS12 4SD             E-Mail: [e-mail]  |
|  England                                                  |
-------------------------------------------------------------
-----------------------------
From:          <[e-mail]>

    Hello Swamy,

a colleague has done examinations on the message
startup time for the PEACE operating system on
the SUPRENUM distributed memory machine.
The paper is called:
    "Overcoming the Startup Time Problem in
     Distributed Memory Architectures"
by
    W. Schroeder-Preikschat from Gmd First in
    Berlin, Germany.

You can obtain it from "Proceedings of the 24th
Hawaii International Conference on System Sciences",
Vol. 1, pp. 551-559, 1991. Additional information on
PEACE or SUPRENUM can be obtained by anonymous ftp at
    ftp.gmd.de in the directory gmd/peace

Cheers, Joerg
--
Joerg Cordsen            | INTERNET:    [e-mail]
c/o GMD-First  Berlin    | KOMEX:       [e-mail]
O-1199 Berlin-Adlershof  | TEL:         [phone]
Rudower Chausee 5 (13.7) | FAX:         [phone]
---------------------------
From:          <[e-mail]>

I'm interested in what you find out. I've been using numbers of 2ms for
workstation RPC and a very *round* number of 100us for MPP message
overhead. The 100us could be very wrong.
-------------------------
From:          John Salmon <[e-mail]>

Rik Littlefield from PNL ([e-mail]) has done a lot of work
unraveling the mysteries of communication timing on the Delta.  The
situation is far from simple.  I know he has given a couple of talks
on the subject.  I recommend contacting him to see if he's got anything
written up.

By the way, you're asking the right question.  Vendors will try to tell
you some hardware number that is only distantly related to the "real" cost
to go from user-code to user-code.

John
 ------------------------------------------------------------------------
From:          <[e-mail]>
To:            <[e-mail]>
>Organization: Minnesota Supercomputer Center, Inc.


Injecting a message into the CM-5 data network takes just over one
microsecond.   On a relatively uncongested communication pattern
it will reach the other end in another microsecond or so.

Note 'startup time' cannot be directly measured, only indirectly, by
comparing the time of an n unit message with an n+1 unit message.  This
function is not linear for all n, however.

<This should include all the operating system or runtime system overheads
<in initiating the communication in addition to hardware setup time.

TMC did the Right Thing and kept the operating system totally out the
picture.  The communication FIFOs are directly mapped into user memory
space so there is no "system" overhead for communication.   (I wish
they went further mapped the FIFOs into the CPU registers themselves,
a la the J machine.)

I'm glad you posted your request.  I think too many people emphasize
bandwidth and forget that low latency is equally important.

--
Alan E. Klietz
Minnesota Supercomputer Center, Inc.
1200 Washington Avenue South
Minneapolis, MN  55415
Ph: [phone]           Internet: [e-mail]