Fast path recommendation for Pea0 on 8.4 blade cluster

[Mike R Smith]

I am trying to determine the best place to set the fastpath for PEA0 traffic on the nodes in an 8.4 blade Integrity cluster. I have reviewed the HP VMS I/O Users Reference Manual. Besides being off the primary cpu, no real clear advantage to any other cpu that doesn't already have something else fast pathed to it. At least none that I see so far.

Any best practices docs or recommendations specifically on this out there? Any pointers or suggestions are appreciated.

[Hoff]

Sure. Pick a core that isn't overloaded. Use T4 or other tools to monitor system and network and clustering activity and (if and when issues surface) tweak the configuration as needed. Done.

[NilakantanM · Accepted solution]

For LAN based clusters
--------------------
The recommended method is to affinitize PEA0 device on to the same CPU where the LAN device used for SCS (Cluster) traffic is affinitized.
for example if EWA0 is used for SCS, then both EWA0 and PEA0 should be on same CPU.


If you have multiple LAN devices for SCS device pick one of the LAN device and affinitize PEA0 on to same CPU as that LAN device. It would be fine to affinitize all the SCS LAN devices and PEA0 to the same CPU, however it can saturate the CPU with interrupt load. (T4 can help if such a situation happens).

The reason why we recommend PEA0 and corresponding LAN device on the same device is multiple reasons.

1: It avoids spinlock contention and busy waits.
2:It helps in better CPU cache utilization. In the networking world as packet passes up the various layers, it is always better to optimize the way CPU cache is utilized.
Reduces latency and helps in higher throughput.
3: It also helps to minize the CPU hand-off cost.

Generally BG device is affinitized to a different CPU than PEA0

For IP based clusters
-----------------------
If Cluster over IP (IPCI) is used, we would recommend BG to be on the same CPU as PEA0 and LAN device used for IP communicatio.
This should be optimal as long as CPU is not saturated and yields better latency and throughput.

OpenVMS Engineering conducts benchmark tests and the same is published during bootcamp/Technical update dates.

These aspects were discussed in bootcamp 2010 sessions.

Regards
Nilakantan
OpenVMS Engineering

[Mike R Smith]

Thanks guys for two very good answers.