Red Hat Bugzilla – Bug 446944
Max latency near 1ms encountered while running cyclictest
Last modified: 2008-08-14 16:51:43 EDT
Description of problem: While running cyclictest on various systems (with load in the background), large (>=1ms) latencies can be observed Version-Release number of selected component (if applicable): 2.6.24.4-30+ How reproducible: Sometimes takes a while but in general show up in less than 5 minutes of runtime Steps to Reproduce: 1. Boot RT kernel 2. run: cyclictest -n -p 90 -t 4 3. run: (cd LTPdir; while true; do ./runltp; done) 4. run: (cd kernelsrc; while true; do make clean bzImage modules; done) Actual results: One of the four cyclictest threads will eventually show a latency in the 900us-1ms range. Avg latency on a dual 2GHz Opteron system will stay in the single microseconds: $ sudo cyclictest -n -p 90 -t 4 0.11 0.18 0.14 1/158 14964 T: 0 ( 3419) P:90 I:1000 C: 819087 Min: 4 Act: 5 Avg: 5 Max: 40 T: 1 ( 3420) P:89 I:1500 C: 546059 Min: 4 Act: 6 Avg: 5 Max: 960 T: 2 ( 3421) P:88 I:2000 C: 409544 Min: 4 Act: 4 Avg: 5 Max: 32 T: 3 ( 3422) P:87 I:2500 C: 327635 Min: 4 Act: 5 Avg: 5 Max: 30 Expected results: No Max Latency over 100us Additional info: I've run this test on a number of the test boxes in Huntsville and all have exhibited this behavior, with production kernel-rt packages.
Clark, try this: # sysctl kernel.sched_nr_migrate=2 And see if that solves you problem. With the new 2.6.24.7-rt7 coming out (with latency stealing), I was hitting that 900+ all the time. Looking at ftrace, I found that the load balancing was taking up to 400us at a time. Compounded by 4way or more boxes that can be running this balancing at the same time. My trace showed that the double_rq_lock took 424us!!! during this balance code and that was because the CPU of the rq it was trying to grab was also trying to do balancing. This needs to be noted (Lana ;-)
Nevermind, I just hit the 900us+ latency with sched_nr_migrate=2. I'll look deeper into it.
What ftrace setup are you using to try and trap this thing? If you can post a setup here, I'll try and duplicate it and get a trace for you.
elevating to blocker status
Created attachment 306803 [details] allow delay to preempt Seems that x86 delay uses tsc, and Andrew Morton made the delay non-preempt to keep problems from non-synced tsc's from causing problems on SMP. This disabling of preemption caused us large latencies. The patch breaks this preemption and removes the 1ms latencies that we are seeing.
Changing status to MODIFIED to show that this bug is considered resolved, but needs testing. The -rt11 version will contain this patch.
kernel-rt-2.6.24.7-60.el5rt was built with -rt11 and is currently under test. Early signs show that the __delay fix has addressed this bug.
fixed in GA
Need to reopen and set state to Modified for 1.0.1 release.
Closing. Fixed in GA.