TCPIP_speed vs DecNet_over_IP_speed

[Sergey Akimov]

I rewrote application (initially writted for DecNet) for working over TCPIP (using TCPIP examples as sample of TCP-sockets programming with SYS$QIO). So, when I tests DecNet (over IP) and pure-IP versions, I see amazing fact: DecNet_over_IP networking work more faster (at least 80% difference) then pure_IP_networking. WHY???
Experiments with IP window size or IP packets size only decrease speed (compared with standard parameters values) :))
I have only one idea - PWIP driver working not over SYS$QIO services, but uses more down level APIs... Isn`t it? If not, why this situation appears? And how to increase IP-netwoking perfomance with SYS$QIO service?

[Ian Miller.]

PWIP does use a lower level interface into the ethernet drivers but that is unlikely to be relvant.

DECnet basically provides a record oriented service (like the native VMS file system) rather than the byte stream service provied by TCP (like native unix file system).

When converting to use TCP rather than DECnet I found that I had to do more work to do the same things (this may be because I'm a VMS programmer not a unix programmer and therefore am not so familier with TCP socket networking).

What I mean is that the performance difference may be due to changes you did to get you application to work over IP.

[Robert Gezelter]

Sergey,

There are many possibilities. What is CPU utilization on the processor running the application).

Another possibility, admittedly an interesting one, is that you are seeing the advantages of using a protocol which has records as an inherant property rather than the streams used in TCP. It is an interesting question as to the exact details of what is happening and where the performance is disappearing.

I hope that the above is helpful.

- Bob Gezelter, http://www.rlgsc.com

[Tim Hughes_3]

Can you measure the actual running packet rate and bytes per packet? Also make sure the DNS lookup is not a problem. DECNET has a name cache but pure IP might do a DNS lookup. Depending on how the "paths" are defined it may take a few tries (with timeouts) to get a match from the DNS server,

Tim

[Martin Kirby_1]

Another possibility, that I have fallen over, is the Nagle algroithm - it depends what you are doing and measuring.

You could try setting TCP_NODELAY on the socket.