On 9/26/18 12:53 PM, John Kacur wrote:
>
>
> On Thu, 13 Sep 2018, Ralf Ramsauer wrote:
>
>> If the histogram output is enabled, cyclictest dumps lines of all latencies,
>> even if the specific latency was never hit. For analysing a histogram, it is
>> sufficient to only dump latencies that actually occured. This patch skips
>> latencies that were never hit with respect to all threads.
>>
>> If cyclictest is used in combination with --nsecs, this removes a lot of
>> output noise.
>>
>> It shrinks output to:
>> $ cyclictest --threads 2 --affinity 2-3 --priority 99 --histofall 100 --loops 100000
>> # /dev/cpu_dma_latency set to 0us
>> policy: fifo: loadavg: 0.03 0.13 0.07 1/560 1248
>>
>> T: 0 ( 1228) P:99 I:1000 C: 100000 Min: 2 Act: 2 Avg: 2 Max: 4
>> T: 1 ( 1229) P:99 I:1000 C: 100000 Min: 2 Act: 2 Avg: 2 Max: 4
>> # Histogram
>> 000002 069425 067138 136563
>> 000003 030536 032850 063386
>> 000004 000039 000012 000051
>> # Total: 000100000 000100000 000200000
>> # Min Latencies: 00002 00002
>> # Avg Latencies: 00002 00002
>> # Max Latencies: 00004 00004 00004
>> # Histogram Overflows: 00000 00000 00000
>> # Histogram Overflow at cycle number:
>> # Thread 0:
>> # Thread 1:
>>
>> Signed-off-by: Ralf Ramsauer <ralf.ramsauer@xxxxxxxxxxxxxxxxx>
>> ---
>> src/cyclictest/cyclictest.c | 13 ++++++++++---
>> 1 file changed, 10 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/cyclictest/cyclictest.c b/src/cyclictest/cyclictest.c
>> index 8d9ec80..0e8459c 100644
>> --- a/src/cyclictest/cyclictest.c
>> +++ b/src/cyclictest/cyclictest.c
>> @@ -1988,16 +1988,23 @@ static void print_hist(struct thread_param *par[], int nthreads)
>> fprintf(fd, "# Histogram\n");
>> for (i = 0; i < histogram; i++) {
>> unsigned long long int allthreads = 0;
>> + unsigned long curr_latency = 0;
>> +
>> + for (j = 0; j < nthreads ; j++) {
>> + curr_latency = par[j]->stats->hist_array[i];
>> + allthreads += curr_latency;
>> + log_entries[j] += curr_latency;
>> + }
>> + if (!allthreads)
>> + continue;
>>
>> fprintf(fd, "%06d ", i);
>>
>> for (j = 0; j < nthreads; j++) {
>> - unsigned long curr_latency=par[j]->stats->hist_array[i];
>> + curr_latency = par[j]->stats->hist_array[i];
>> fprintf(fd, "%06lu", curr_latency);
>> if (j < nthreads - 1)
>> fprintf(fd, "\t");
>> - log_entries[j] += curr_latency;
>> - allthreads += curr_latency;
>> }
>> if (histofall && nthreads > 1) {
>> fprintf(fd, "\t%06llu", allthreads);
>> --
>> 2.19.0
>>
>>
>
> Note, although I initially accepted this patch I may have to revert it.
> The output of the histogram is not consumed just by humans but by
> software, namely rteval
Narf, didn't know.
>
> I am leaving the patch in for now and considering whether rteval could be
> modified, but this is a heads-up.
>
> The other option, is for this patch to be an option
Give me a couple of days and let me have a look at rteval, zero-lines
are really useless. Maybe there's a way for rteval to support any output
format of cyclictest.
Ralf
>
> Thanks
>
> John
>
