SQL*Net performance - why this way?

Dear Netters,

I have done some performance measurements on using Oracle SQL*Forms in client/server mode. The client was always a PC (386SX/16MHz to 486/50MHz) and was used against a remote database server.

The results were (using SQL*Forms 3.0 and 386SX 16 MHz IBM PS/2 55SX)

  Simple query (13 different SQL statement executions)

  Time Relative Server/Protocol

  3.5s  1.00      IBM 3090 VM/XA         APPC
  3.3s  0.94      HP9000/817 HP-UX       TCP/IP
  3.2s  0.91      IBM RS6000 340 AIX     TCP/IP
  2.6s  0.74      IBM PS/2 70 OS/2 1.3   NetBIOS
  2.6s  0.74      Tricord Novell NetWare SPX/IPX

  Complex update

  Time Relative Server/Protocol

  20.9s 1.00      IBM 3090 VM/XA         APPC
  17.5s 0.84      HP9000/817 HP-UX       TCP/IP
  17.2s 0.82      IBM RS6000 340 AIX     TCP/IP
  13.1s 0.63      Tricord Novell NetWare SPX/IPX

Using 486 25 MHz client (Dell 425DE) the results for the complex update were:

  Time Relative Server/Protocol

  14.2s 1.00      IBM 3090 VM/XA         APPC
   6.1s 0.43      IBM RS6000 340 AIX     TCP/IP
   5.5s 0.39      Tricord Novell NetWare SPX/IPX

All tests were performed without other users whenever possible. We used 4Mbps Token Ring in all tests. The time was taken when repeating the test during the same session (the first timing during a session is about 30..50% slower than subsequent timings).

Could anybody explain WHY we got those results: Why SPX/IPX and NetBIOS were faster than TCP/IP, particularly on the 386SX?

Are there any parameters which could be tuned to gain better performance? We used HP ARPA 2.11 TCP/IP package on the PCs. I tried to tune some parameters without any success.

Do the results match any other measurements?

Why 386SX is so much slower than 486 25 MHz?

By the way, the complex update takes about 2 seconds when run on the UNIX servers without SQL*Net or using a powerful UNIX workstation as a client.

--
Juhani Jaakola
jaakola_at_cc.helsinki.fi