[dpdk-dev] [PATCH] app/test-compress-perf: report header improvement

Trybula, ArturX arturx.trybula at intel.com
Thu Jun 27 13:46:39 CEST 2019



-----Original Message-----
From: Shally Verma [mailto:shallyv at marvell.com] 
Sent: Thursday, June 27, 2019 13:01
To: Trybula, ArturX <arturx.trybula at intel.com>; dev at dpdk.org; Trahe, Fiona <fiona.trahe at intel.com>; Dybkowski, AdamX <adamx.dybkowski at intel.com>
Subject: RE: [dpdk-dev] [PATCH] app/test-compress-perf: report header improvement



> -----Original Message-----
> From: Trybula, ArturX <arturx.trybula at intel.com>
> Sent: Thursday, June 27, 2019 2:29 PM
> To: Shally Verma <shallyv at marvell.com>; dev at dpdk.org; Trahe, Fiona 
> <fiona.trahe at intel.com>; Dybkowski, AdamX <adamx.dybkowski at intel.com>
> Subject: [EXT] RE: [dpdk-dev] [PATCH] app/test-compress-perf: report 
> header improvement
> 
> External Email
> 
> ----------------------------------------------------------------------
> 
> 
> -----Original Message-----
> From: Shally Verma [mailto:shallyv at marvell.com]
> Sent: Wednesday, June 26, 2019 19:04
> To: Trybula, ArturX <arturx.trybula at intel.com>; dev at dpdk.org; Trahe, 
> Fiona <fiona.trahe at intel.com>; Dybkowski, AdamX 
> <adamx.dybkowski at intel.com>
> Subject: RE: [dpdk-dev] [PATCH] app/test-compress-perf: report header 
> improvement
> 
> 
> 
> > -----Original Message-----
> > From: dev <dev-bounces at dpdk.org> On Behalf Of Artur Trybula
> > Sent: Monday, June 3, 2019 3:54 PM
> > To: dev at dpdk.org; fiona.trahe at intel.com; arturx.trybula at intel.com; 
> > adamx.dybkowski at intel.com
> > Subject: [dpdk-dev] [PATCH] app/test-compress-perf: report header 
> > improvement
> >
> > This patch adds extra features to the compress performance test. 
> > Some important parameters (memory allocation, number of ops, number 
> > of
> > segments) are calculated and printed out on the screen.
> >
> > Signed-off-by: Artur Trybula <arturx.trybula at intel.com>
> > ---
> >  app/test-compress-perf/main.c | 105
> > +++++++++++++++++++++++++++++++---
> >  1 file changed, 98 insertions(+), 7 deletions(-)
> >
> ......
> >
> > -
> > +static void
> > +print_report_header(void)
> > +{
> > +	uint32_t opt_total_segs = DIV_CEIL(tests_res.input_data_sz,
> > +			MAX_SEG_SIZE);
> > +
> > +	if (tests_res.total_buffs > 1) {
> > +		printf("\nWarning: for the current input parameters number"
> > +				" of ops is higher than one, which may result"
> > +				" in sub-optimal performance.\n");
> > +		printf("To improve the performance (for the current"
> > +				" input data) following parameters are"
> > +				" suggested:\n");
> > +		printf("	• Segment size: %d\n", MAX_SEG_SIZE);
> > +		printf("	• Number of segments: %u\n", opt_total_segs);
> > +	} else if (tests_res.total_buffs == 1) {
> > +		printf("\nWarning: There is only one op with %u segments –"
> > +				" the compression ratio is the best.\n",
> > +			tests_res.segments_per_last_buff);
> > +		if (tests_res.segment_sz < MAX_SEG_SIZE)
> > +			printf("To reduce compression time, please use"
> > +					" bigger segment size: %d.\n",
> > +				MAX_SEG_SIZE);
> > +		else if (tests_res.segment_sz == MAX_SEG_SIZE)
> > +			printf("Segment size is optimal for the best"
> > +					" performance.\n");
> 
> [Shally] Why these kind of warnings? If total_bufs > 1, then, how 
> behavior would change? Same question for segment size as well?
> [Artur] I suppose you mean the second one "Warning: There is only one ...".
> It's just to preserve the structure of the report for both variants.  
> But I agree, there is nothing wrong if there is only one op.
> Please notice that each op is processed independently, in most cases 
> it implies lower compression ratio. The best situation is to have all 
> the input data aggregated into one op (even if partitioned across many mbufs).
> 
[Shally] Then can we better covert this in terms of "stateless" and "stateful" . Example, Say it like: 
Since test app support stateless mode only, thus if input data is segmented across multiple buffers (i.e. op) Then each will be processed independently (i.e. in stateless mode) . Thus to gain better results, it might be more useful to put all data in one buffer.
[Artur] We started from an idea of simple checking memory allocation. From my point of view we should avoid making it too complex unless there is a clear need. This is a "simple tool" for quick verification what is the effect of the input parameters. It's for developers familiar with the compression. I agree with you that in case of supporting both stateless and stateful compression such information should be printed out.


> > +	} else
> > +		printf("Warning: something wrong happened!!\n");
> > +
> > +	printf("\nFor the current input parameters (segment size = %u,"
> > +			" segments number = %u):\n",
> > +		tests_res.segment_sz,
> > +		tests_res.segments_per_buff);
> > +	printf("	• Total number of segments: %d\n",
> > +		tests_res.total_segments);
> > +	printf("	• %u segments %u bytes long, last segment %u"
> > +			" byte(s) long\n",
> > +		tests_res.total_segments - 1,
> > +		tests_res.segment_sz,
> > +		tests_res.last_segment_sz);
> > +	printf("	• Number of ops: %u\n", tests_res.total_buffs);
> > +	printf("	• Total memory allocation: %u\n",
> > +		(tests_res.total_segments - 1) * tests_res.segment_sz
> > +		+ tests_res.last_segment_sz);
> > +	if (tests_res.total_buffs > 1)
> > +		printf("	• %u ops: %u segments in each,"
> > +				" segment size %u\n",
> > +			tests_res.total_buffs - 1,
> > +			tests_res.segments_per_buff,
> > +			tests_res.segment_sz);
> > +	if (tests_res.segments_per_last_buff > 1) {
> > +		printf("	• 1 op %u segments:\n",
> > +				tests_res.segments_per_last_buff);
> > +		printf("		o %u segment size %u\n",
> > +			tests_res.segments_per_last_buff - 1,
> > +			tests_res.segment_sz);
> > +		printf("		o last segment size %u\n",
> > +			tests_res.last_segment_sz);
> > +	} else if (tests_res.segments_per_last_buff == 1) {
> > +		printf("	• 1 op (the last one): %u segment %u"
> > +				" byte(s) long\n\n",
> > +			tests_res.segments_per_last_buff,
> > +			tests_res.last_segment_sz);
> > +	}
> > +}
> >
> >  int
> >  main(int argc, char **argv)
> > @@ -533,8 +622,9 @@ main(int argc, char **argv)
> >  	else
> >  		level = test_data->level.list[0];
> >
> > +	print_report_header();
> > +
> [Shally] looks like we're printing input characteristics and possible 
> performance behavior. Is that the intention of this API?
> [Artur] That was the idea to have a tool for verification how the data 
> partitioning affects the compression performance. Your description 
> Shally is very accurate: "... printing input characteristics and 
> possible performance behavior" and that is the intention of this API.
[Shally] Then probably we can add this in function description.
[Artur] To be precise it is not a typical API function. It's an internal (static) function of the perf test. 
> 
> 
> >  	printf("Burst size = %u\n", test_data->burst_sz);
> > -	printf("File size = %zu\n", test_data->input_data_sz);
> >
> >  	printf("%6s%12s%17s%19s%21s%15s%21s%23s%16s\n",
> >  		"Level", "Comp size", "Comp ratio [%]", @@ -612,3 +702,4
> @@
> > main(int argc, char **argv)
> >  	}
> >  	return ret;
> >  }
> > +
> > --
> > 2.17.1



More information about the dev mailing list