Release of TCA Version 7.9

Summary:
========

Release 7.9 is an optimization release.  No added functionality.  The
system just runs a lot faster, in some cases, faster than TCX.

For comparison, I have run a number of tests comparing TCA 7.8, TCA 7.9
and TCX 12.3 (The most recent TCX version I can find).

Test 1:
=======

One process makes 200 queries of another process.  The data size of each
request message varies from zero bytes to 1 Megabyte.  The reply is
alway one integer (4 bytes).

All tests were conducted on a sparc 10 running Sunos.  All processes
where run on the same machine.  Using different machines runs up against
the limited bandwidth of the network (10 Mbits/s).

                 TCA 7.8               TCX 12.3              TCA 7.9
Bytes/msg    msg/s     bytes/s     msg/s     bytes/s     msg/s     bytes/s
---------    -----------------     -----------------     -----------------
   0/4*      242          0.00     562      2,249.84     400          0.00
    16       233      3,735.20     545      8,722.55     370      5,921.10
   256       221     56,501.66     508    130,134.53     352     90,226.62
    1K       209    213,619.94     472    483,341.50     333    341,455.72
    4K       119    485,990.09     256  1,046,894.31     264  1,081,642.00
   16K        53    871,264.56     105  1,722,838.00     134  2,188,539.75
   64K        17  1,118,213.75      31  2,000,986.88      42  2,735,001.75
  256K         5  1,194,023.12       9  2,237,433.00      12  3,059,980.25
    1M         1  1,220,855.00       2  2,257,506.00       3  3,171,188.75

* TCX version crashed when a null format was used, so I used a single
integer.

Comments: For TCA, the maximum message rate has almost doubled and the
maximum through put has almost tripled.  Comparing TCA and TCX, we see
that for very small messages, TCX is faster by 40%. For larger messages,
TCA has a higher throughput.  For 1Megabyte messages, TCA is 50% faster
than TCX.  Somewhere around 2K, is the cross over point where the
systems have the same message rate and throughput.

Test 2:
=======

Send 200 messages from one process to another, optimizing the program
for raw speed using each messaging system. 

                 TCA 7.8               TCX 12.3              TCA 7.9
Bytes/msg    msg/s     bytes/s     msg/s     bytes/s     msg/s     bytes/s
---------    -----------------     -----------------     -----------------
   0/4*      412          0.00     2578     10,311.94    657          0.00
    16       363      5,801.64     2170     34,715.44    582      9,317.85
   256       327     83,705.95     1414    361,880.94    426    108,941.50
    1K       280    286,760.03      853    873,005.00    386    394,999.22
    4K       143    585,518.81      258  1,057,488.00    302  1,236,209.12
   16K        58    955,469.81      104  1,699,761.38    135  2,210,402.50
   64K        18  1,156,581.38       29  1,913,896.88     42  2,775,157.00
  256K         5  1,218,955.25        8  2,013,264.62     12  3,068,828.50
    1M         1  1,229,735.12        2  2,140,054.50      3  3,199,656.00

* TCX version crashed when a null format was used, so I used a single
integer.

Comments: The TCA maximum message rate has increased by 60% and the
through put has almost tripled.  Comparing TCA and TCX, we again see the
same pattern where TCX is faster for messages will little data and TCA
is faster for larger messages.  This is mostly due to the topology.  TCX
messages need only make a single hop, while TCA messages go through the
central server.  The TCX programs were also optimized to block on the
read port and not do a select to see which socket/file had available
input.  Doing a select would slow the message rate down and is needed if
the module receives input from more than one other module.

Test 3:
=======

Same as test 2, except using the central routing feature of TCX.  This
was a recent addition that allows messages to be routed through the
central server.

                 TCA 7.8               TCX 12.3              TCA 7.9
Bytes/msg    msg/s     bytes/s     msg/s     bytes/s     msg/s     bytes/s
---------    -----------------     -----------------     -----------------
   0/4*      412          0.00     657      2,627.23     657          0.00
    16       363      5,801.64     507      8,111.76     582      9,317.85
   256       327     83,705.95     509    130,202.04     426    108,941.50
    1K       280    286,760.03     336    344,414.69     386    394,999.22
    4K       143    585,518.81     167    685,426.12     302  1,236,209.12
   16K        58    955,469.81      58    950,648.81     135  2,210,402.50
   64K        18  1,156,581.38      17  1,083,263.25      42  2,775,157.00
  256K         5  1,218,955.25       4  1,160,282.50      12  3,068,828.50
    1M         1  1,229,735.12       1  1,202,580.62       3  3,199,656.00

Comments:  When routing the messages through the central server, TCX
slows down considerably.  The message rate for small messages is about
the same.  (Don't read too much into the exact match for the smallest
messages.)  For larger messages, the rate is about the same for TCA 7.8.

Conclusions:
============

If you need a very high message rate with little data to be transfered,
then TCX has some speed advantages.  For significant data, or if
synchronization is needed, the TCX speed advantages disappear, and for
large messages, TCA 7.9 is the clear winner.