11

I'm trying to understand some data that has been pulled from SAR. I have three main questions about this. Ultimately, I'd like to determine how many CPUs were idle at each sampling interval across a cluster of servers.

  1. Many of the CPUs are not showing up in every entry. Is this expected and What exactly does that mean? Is it related to #2?
  2. There are unused lines (CPU = U). The documentation says "U indicates the system-wide Unused capacity". I can't find a precise definition of "system-wide Unused capacity" or any definition at all, really. I'm not sure how to interpret a line that tells something like "the unused capacity was 70% idle."
  3. Lastly, I'm unsure of how the - or all line is calculated. I would think it's the average of all the CPUs but when I do the math across all CPUs, I get a vastly different answer than what is on that line. Can anyone tell me exactly what goes into that calculation? Looking closely at this related question about SAR it appears that the system-wide idle percentage is the sum of the product of each CPU's idle percentage and the 'physc' value. Unfortunately, I don't have the physc or entc% (assuming there is one) so I can't verify this with my own data. If that's correct, does it mean I need the physc values to truly understand idle percentage?

Here are some of examples of what I'm seeing. These are all from the same day.

CPU | Idle    CPU | Idle    CPU | Idle
----------    ----------    ----------
0   | 8       0   | 15      0   | 17
1   | 25      1   | 94      1   | 32
2   | 79      2   | 100     2   | 97
3   | 62      3   | 99      3   | 71
4   | 5       4   | 13      4   | 5
5   | 7       5   | 13      5   | 23
6   | 6       6   | 99      6   | 71
7   | 7       7   | 44      7   | 98
8   | 11      8   | 12      8   | 48
9   | 17      12  | 0       12  | 38
10  | 33      16  | 12      16  | 37
11  | 64      20  | 3       20  | 42
12  | 6        U  | 95       U  | 97
13  | 6        -  | 15       -  | 85
14  | 6       
15  | 6       
16  | 12      
17  | 15      
18  | 62      
19  | 69      
20  | 7       
21  | 7       
22  | 6       
23  | 7       
 U  | 80       
 -  | 15       

case 1: avg(24): 22
case 2: avg(12): 42
case 3: avg(12): 48

This data is produced by a script that runs: sar -P ALL 1 1 It then runs an awk command. I'm not good with awk but these are clearly the important parts:

Filter: /System|AIX|^$|%/ {next}

Parse: {k=0;if(NR==7) k=1} {sub("^-", "all", $1); cpu=$(1+k); user=$(2+k); sys=$(3+k); io=$(4+k); idle=$(5+k)}

This seems correct based on what little I understand of awk and what I see from examples of the output.

If I assume that the missing values are all zero for case 2, the average is 21 which seems somewhat consistent with case 1. However, if I make that assumption for case 3, I get 24% which is completely at odds with the 85% percent value given by sar for the overall CPU idle.

Here's a graph of a full day's captures (every 30 seconds):

enter image description here

When there is very little 'system-wide' idle time, the correlation between the average CPU idle and the 'system-wide' idle is almost perfect. But as the 'system-wide' idle time increases, the correlation becomes much weaker. Working on the assumption that these are deterministic machines, that tells me that the data I have is not giving the full picture. But how much do I care?

I don't fully understand why some CPUs are not being reported at each point but the ones that are missing are not evenly distributed as seen in the examples above. Also from reading this redbook, I take it that these must be logical CPUs and that without the physc numbers, I think there's not much I can do with these values. I've tried to use the U value in various equations but I haven't found anything sensible. It's not even clear to me that the overall idle percentage can be taken at face value.

NOTE: There is something wrong with the capture of this data from sar is a completely valid answer for #1, if it's the case it should always return.

8
  • What command are you running to get that output? It doesn't look like the standard sar -P ALL output.
    – Swiss
    Mar 6, 2017 at 23:22
  • @Swiss This is coming from a script that I didn't write. It calls sar -P ALL 1 1 and then uses awk to break out the cpu number and then the user, system,IO-wait, and idle percentages. I'll add more info to your answer.
    – JimmyJames
    Mar 7, 2017 at 14:14
  • @Swiss I edited the example to be more reflective of what the output of the script looks like.
    – JimmyJames
    Mar 7, 2017 at 14:33
  • could you provide the output of sar -P ALL directly, rather than the output of this script? It's a non-standard script and no one can tell you what it does without seeing it.
    – Swiss
    Mar 7, 2017 at 16:31
  • @Swiss unfortunately I can't. I can see the script and it's output. Is there something missing from the description of what it does that I can clarify?
    – JimmyJames
    Mar 7, 2017 at 16:34

1 Answer 1

4
+50

The output you have provided looks different from the standard sar -P ALL or sar -u output. I'm not sure if you hand formatted it, or if you're running it through another tool, but I think there's enough information there to figure this out.

Here's the important piece of information, obtained from the man page for sar

Note: On SMP machines a processor that does not have any activity at all (0.00 for every field) is a disabled (offline) processor.

Since you are running in a cluster, it seems fairly safe to assume that you are using SMP machines.

Note that in example 2 and 3, only 12 of the 24 cores are reporting statistics. If you assume that these cores are disabled, as mentioned in the man page, then the statistics make sense.

Let's update your data as follows, to indicate a disabled core with -

0   | 8       0   | 15      0   | 17
1   | 25      1   | 94      1   | 32
2   | 79      2   | 100     2   | 97
3   | 62      3   | 99      3   | 71
4   | 5       4   | 13      4   | 5
5   | 7       5   | 13      5   | 23
6   | 6       6   | 99      6   | 71
7   | 7       7   | 44      7   | 98
8   | 11      8   | 12      8   | 48
9   | 17      9   | -       9   | -
10  | 33      10  | -       10  | -
11  | 64      11  | -       11  | -
12  | 6       12  | 0       12  | 38
13  | 6       13  | -       13  | -
14  | 6       14  | -       14  | -
15  | 6       15  | -       15  | -
16  | 12      16  | 12      16  | 37
17  | 15      17  | -       17  | -
18  | 62      18  | -       18  | -
19  | 69      19  | -       19  | -
20  | 7       20  | 3       20  | 42
21  | 7       21  | -       21  | -
22  | 6       22  | -       22  | -
23  | 7       23  | -       23  | -

Then we can use the following to calculate the averages (this is a quick oneliner I wrote, I'm sure something better could be written.)

$ awk '{idle = $3; output += idle; if (idle >= 0) {cores += 1} } END {printf "Ave idle%%: %f  Cores: %d\n", output / cores, cores }' input_file
Ave idle%: 22.208333  Cores: 24
$ awk '{idle = $6; output += idle; if (idle >= 0) {cores += 1} } END {printf "Ave idle%%: %f  Cores: %d\n", output / cores, cores }' input_file
Ave idle%: 42.000000  Cores: 12
$ awk '{idle = $9; output += idle; if (idle >= 0) {cores += 1} } END {printf "Ave idle%%: %f  Cores: %d\n", output / cores, cores }' input_file
Ave idle%: 48.250000  Cores: 12

Note that the number of cores in example 2 and 3 is 12, and the averages match what you see in your example output.

It looks like at some point between your first and second case, half your CPU cores have been disabled.

A quick summary on your questions:

  1. Missing data probably means that processor core is disabled.
  2. The U in the average line is different than the U in that man page. The U referred to in the man page should show up under the processor ID column.
  3. The output you've provided is different than the standard sar output and there's not enough information provided to determine what the U or all on the average line refer to. The first number appears to be the idle% on active cores though.
4
  • "Note that the number of cores in example 2 and 3 is 12, and the averages match what you see in your example output." Sorry,I may not have been clear, I calculated the average from the numbers. That's not coming from sar. The U and all values are coming from sar as lines. I broke them out in my answer since they fundamentally different things than the cpu values, from what I understand.
    – JimmyJames
    Mar 7, 2017 at 14:17
  • If the cores are disabled, why is the 'system-wide' idle time so much less in case 2 than in case 3? I'm struggling to understand how they can have nearly the same average idle time across the same 12 cpus and yet the system-wide idle time be completely divergent.
    – JimmyJames
    Mar 7, 2017 at 14:21
  • I do see that when machines are under constant load, I do see output from every processor. This would suggest to me that you are correct that these are disabled. I'm just struggling with how the system wide idle relates to the idle of the non-disabled cpus.
    – JimmyJames
    Mar 7, 2017 at 14:39
  • I updated my question with a few changes to clarify the problem. So looking at case 2, if 12 cpus are disabled, and the sys-idle percentage were based on all 24, the idle% could not go below 50%. So given that I have 15% that means the sys idle% must be ignoring disabled CPUs, right?
    – JimmyJames
    Mar 7, 2017 at 16:32

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .