Low Latency Performance Tuning for Red Hat Enterprise Linux 7

Re:Low Latency Performance Tuning for Red Hat Enterprise Lin... by smm_latency on Wednesday February 11, 2015 @03:30PM (#9035519)

From my experience, it's very difficult to achieve low latency on modern Intel processors (after Sandy Bridge) because of SMI interrupts:

https://en.wikipedia.org/wiki/System_Management_Mode

Some of the SMI interrupts can be disabled by smictrl, but there usually remains an interrupt every 10-20 seconds with 100+ microsecond latency. See plot here:

http://wiki.linuxcnc.org/cgi-bin/wiki.pl?FixingSMIIssues

SMI interrupts are used for fan / thermal region control and cannot be fully disabled.

[ Parent | on HN ]

Re:Low Latency Performance Tuning for Red Hat Enterprise Lin... by jeremyeder on Wednesday February 11, 2015 @03:53PM (#9035639)

"cannot be fully disabled"

...depends on the gear. SMI used to be (4-5 years ago) a much larger problem than they are now.

I agree with you in that context, and it's why so few systems are certified for Red Hat's Realtime kernel. They are simply not all created equal.

But I'd encourage you to review the results of any of the 25+ benchmarks we did with STAC over the last few years.

We didn't see much (if any) SMI interference on the gear we had, which was off the shelf regular servers, with WSM, SNB, IVB and HSW. All the hardware, software and config is disclosed within those benchmark write-ups.

There is some tooling called hwlat that can detect and report SMIs. It's in the rt-tests package.

Happy tuning!

[ Parent | on HN ]