[PATCH] app/dma-perf: add average latency per worker

[fengchengwen]

Hi Vipin,

On 2024/2/27 17:50, Varghese, Vipin wrote:
> 
> On 2/23/2024 3:15 PM, fengchengwen wrote:
>> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>>
>>
>> Hi Vipin,
>>
>> On 2023/12/20 0:40, Vipin Varghese wrote:
>>> Modify the user display data with total average latency per worker.
>>>
>>> Signed-off-by: Vipin Varghese <vipin.varghese at amd.com>
>>> ---
>>>   app/test-dma-perf/benchmark.c | 3 ++-
>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/app/test-dma-perf/benchmark.c b/app/test-dma-perf/benchmark.c
>>> index 9b1f58c78c..8b6886af62 100644
>>> --- a/app/test-dma-perf/benchmark.c
>>> +++ b/app/test-dma-perf/benchmark.c
>>> @@ -470,7 +470,8 @@ mem_copy_benchmark(struct test_configure *cfg, bool is_dma)
>>>                bandwidth_total += bandwidth;
>>>                avg_cycles_total += avg_cycles;
>>>        }
>>> -     printf("\nTotal Bandwidth: %.3lf Gbps, Total MOps: %.3lf\n", bandwidth_total, mops_total);
>>> +     printf("\nAverage Cycles/op: %.2lf, Total Bandwidth: %.3lf Gbps, Total MOps: %.3lf\n",
>>> +                     (float) avg_cycles_total / nb_workers, bandwidth_total, mops_total);
> 
> thanks for the suggestions, please find my observations below
> 
>> Because this is total stats, suggest add Total prefix, e.g. "Total Average Cycles/op"
> I did not follow this, so please let me try to explain my understanding. For `n` operation we count the average cycles, then
> we add the cycles to form `total average cycles`; this is then divide by `n` operations. Making this per operation what is the
> average cycles taken for the round trip time. Hence `Total Average Cyeles/op` does not sound right, but `Average Cycles / op` does.

OK
My intention was to differentiate (since they have the same beginning).

>> I think print format keep one-digit precision is enough. Also please modify CSV_TOTAL_LINE_FMT
>> make sure the csv also have same precision of Cycles/op.
> 
> We have checked the CSV formatting a find
> 1. the precision for average cycle/op is 2 digits precision.
> 2. already the CVS format has average cycles/op integrated.
> 
> Hence no change is required.

It's weird.

We can see more clear when add together:
#define CSV_TOTAL_LINE_FMT "Scenario %u Summary, , , , , ,%u,%.2lf,%u,%.3lf,%.3lf\n"
	snprintf(output_str[MAX_WORKER_NB], MAX_OUTPUT_STR_LEN, CSV_TOTAL_LINE_FMT,
			cfg->scenario_id, nr_buf, memory * nb_workers,
			avg_cycles_total / nb_workers, bandwidth_total, mops_total);
The bandwidth_total, mops_total both are float, which take last two "%.31f", then "%u" is for "avg_cycles_total / nb_workers"

Also, avg_cycles_total / nb_workers both are u32, and its result is u32 if not with force convert.

You could modify with "avg_cycles_total*1.0 / nb_workers", then compile will output warning:

[2035/3727] Compiling C object app/dpdk-test-dma-perf.p/test-dma-perf_benchmark.c.o
../../dpdk/app/test-dma-perf/benchmark.c: In function ‘mem_copy_benchmark’:
../../dpdk/app/test-dma-perf/benchmark.c:26:28: warning: format ‘%u’ expects argument of type ‘unsigned int’, but argument 7 has type ‘double’ [-Wformat=]
 #define CSV_TOTAL_LINE_FMT "Scenario %u Summary, , , , , ,%u,%.2lf,%u,%.3lf,%.3lf\n"

Thanks
                            ^
> 
>>
>> Thanks
>>
>>>        snprintf(output_str[MAX_WORKER_NB], MAX_OUTPUT_STR_LEN, CSV_TOTAL_LINE_FMT,
>>>                        cfg->scenario_id, nr_buf, memory * nb_workers,
>>>                        avg_cycles_total / nb_workers, bandwidth_total, mops_total);
>>>
> .