Understanding Linux CPU stats

http://blog.scoutapp.com/articles/2015/02/24/understanding-linuxs-cpu-stats

by itsderek23 • 11 years ago

33 7 11 years ago

blog.scoutapp.com

etep• 11 years ago

I would like to see stats beyond those provided by top. For example, memory bandwidth (broken into read and write bandwidth). IO bandwidth. CPU cache hit rates. Instructions per cycle. Energy consumption. AFAIK there is no really nice way to see all the true under the hood perf stats. Yes I know about perf, LIKWID, and other similar tools -- point being, I think there is an opportunity to put this kind of monitoring into a tool at the level of scout and truly have something informative.

suprjami• 11 years ago

You should look into PCP.

If it's exposed to userspace, you can log it. You can then replay it or graph it, changing the granularity from the sample period upwards.

bnegreve• 11 years ago

You can check perf tool.

E.g.

$> perf stat echo "hi"

     Performance counter stats for 'echo hi':

              0.718209      task-clock (msec)         #    0.595 CPUs utilized          
                     2      context-switches          #    0.003 M/sec                  
                     0      cpu-migrations            #    0.000 K/sec                  
                    59      page-faults               #    0.082 M/sec                  
               844,067      cycles                    #    1.175 GHz                    
               561,879      stalled-cycles-frontend   #   66.57% frontend cycles idle   
       <not supported>      stalled-cycles-backend   
               611,918      instructions              #    0.72  insns per cycle        
                                                      #    0.92  stalled cycles per insn
               127,740      branches                  #  177.859 M/sec                  
                 7,279      branch-misses             #    5.70% of all branches        
    
           0.001207821 seconds time elapsed

For more stats, check perf list

glhaynes• 11 years ago

Where is time spent transitioning from userspace to kernel and back accounted for? And (perhaps relatedly), when does the accounting happen? After each timeslice?

nerdcity• 11 years ago

Why not describe each element in the actual order that top actually presents them, from left to right, instead of jumping around? #flow

kagamine• 11 years ago

Flow is nice, but the article does explain in the paragraph for id that us & id should be close to 100% when added. That's why they are placed together in the article.

pmontra• 11 years ago

I confirm that a high waiting time can be due to a damaged disk. I had a server that became slower over the time notwithstanding a constant number of requests. It took me a while to notice a 90% waiting time in top. A disk was dying, I got it replaced and the server got speedy again.

Understanding Linux CPU stats

Related Stories