Re: [dpdk-dev] [PATCH 2/3] app/compress-perf: add performance measurement

["Verma, Shally" <Shally.Verma@cavium.com>]

>-----Original Message-----
>From: Jozwiak, TomaszX <tomaszx.jozwiak@intel.com>
>Sent: 06 November 2018 14:36
>To: Verma, Shally <Shally.Verma@cavium.com>; dev@dpdk.org; Trahe, Fiona <fiona.trahe@intel.com>; akhil.goyal@nxp.com
>Subject: RE: [dpdk-dev] [PATCH 2/3] app/compress-perf: add performance measurement
>
>External Email
>
>> -----Original Message-----
>> From: Verma, Shally [mailto:Shally.Verma@cavium.com]
>> Sent: Tuesday, November 6, 2018 9:16 AM
>> To: Jozwiak, TomaszX <tomaszx.jozwiak@intel.com>; dev@dpdk.org; Trahe,
>> Fiona <fiona.trahe@intel.com>; akhil.goyal@nxp.com
>> Subject: RE: [dpdk-dev] [PATCH 2/3] app/compress-perf: add performance
>> measurement
...

>> >> >> >+
>> >> >> >+       /* Window size */
>> >> >> >+       if (test_data->window_sz != -1) {
>> >> >> >+               if (param_range_check(test_data->window_sz,
>> >> >> >+ &cap->window_size)
>> >> >> What if cap->window_size is 0 i.e. implementation default?
>> >> >
>> >> >TJ: You probably mean cap->window_size.increment = 0 (because
>> >> >cap->window_size is a structure). In that case we check if
>> >> >test_data->window_sz >=min and test_data->window_sz <= max only,
>> >> because increment = 0 means (base on compression API) we have only
>> >> one value of windows_size (no range is supported).
>> >> But PMD can set min and max too 0 for such case.
>> >
>> >TJ: I can't see any issue in that case too. Maybe I don't understand what you
>> mean but the logic is as follow:
>> >1)  if you pass '--window-sz  ...' param. into command line your
>> >intention is to force that value of window size during test. We check is this
>> value is allow (by param_range_check() function).
>> >2) if you plan to use default value - just don't pass '--window-sz'
>> >param. in command line at all. In that case we get windows size from
>> >window_size.max field, so if window_size.min= window_size.max=0
>> test_data->window_sz will be zero, as well.
>> >If you mean that behavior is not good - I will be grateful for other
>> suggestions.
>>
>> This is fine. but I am thinking of 3rd case here:
>> c) user pass window sz but PMD window_sz.min = max = 0, then user
>> requested windowsz is not applicable right?!
>
>In that case - true. There'll be fail :
>"Compress device does not support this window size\n");
>So what is your proposal for  that case?
>
We can set to window size to implementation default and add in diagnostic of used window sz for test run.
No need to fail here I believe.

Thanks
Shally

>
>
>
>
>>
>> >
>> >>
>> >> >
>> >> >
>> >> >
>> >> ....
>> >>
>> >> >> >+
>> >> >> >+               if (fread(data, data_to_read, 1, f) != 1) {
>> >> >> >+                       RTE_LOG(ERR, USER1, "Input file could not be read\n");
>> >> >> >+                       goto err;
>> >> >> >+               }
>> >> >> >+               if (fseek(f, 0, SEEK_SET) != 0) {
>> >> >> >+                       RTE_LOG(ERR, USER1,
>> >> >> >+                               "Size of input could not be calculated\n");
>> >> >> >+                       goto err;
>> >> >> >+               }
>> >> >> >+               remaining_data -= data_to_read;
>> >> >> >+               data += data_to_read;
>> >> >> It looks like it will run 2nd time only if input file size < input
>> >> >> data size in which case it will just keep filling input buffer
>> >> >> with repeated
>> >> data.
>> >> >> Is that the intention here?
>> >> >
>> >> >TJ: Yes exactly. If test_data->input_data_sz is bigger than
>> >> >actual_file_sz then we fill the buffer with repeated data from file
>> >> >to fill
>> >> whole buffer.
>> >> I mentioned in one of the earlier reply, wont that then influence the
>> >> compression behaviour and o/p? my suggestion was to work on actual
>> >> user provided input to take perf to get actual perf for given content.
>> >
>> >TJ: You right, but this solution is flexible. You can pass '
>> >--extended-input-sz" or not, so you can use original input data or extend it
>> if you want.
>> Ok. but still not sure if it's really needed. Might be practically most of the time
>> it wont be exercised. No hard opinion on this though.
>>
>> Thanks
>> Shally
>> >
>> >>
>> >> >
>> >> >>
>> >> ...
>> >>
>> >> >> >+                       if (data_addr == NULL) {
>> >> >> >+                               RTE_LOG(ERR, USER1, "Could not
>> >> >> >+ append data\n");
>> >> >> Since a new buffer per segment is allocated, so is it possible for
>> >> >> append to fail? think, this check is redundant here.
>> >> >
>> >> >TJ: Yes, you're right, it should never fail. But I think it's good
>> >> >coding practice
>> >> to add the check just in case.
>> >> >
>> >> Unless it is called in data path which might cost perf a bit.
>> >
>> >TJ:  prepare_bufs() is out of perf measurement, so shouldn't impact to
>> >measurements. The performance measurement is inside
>> >main_loop() only.
>> >
>> >
>> >Br, Tomek
>> >
>> >>
>> >> Thanks
>> >> Shally
>> >>
>> >> >> >+                               return -1;
>> >> >> >+                       }
>> >> >> >+
>> >> >> >+                       rte_memcpy(data_addr, input_data_ptr, data_sz);
>> >> >> >+                       input_data_ptr += data_sz;
>> >> >> >+                       remaining_data -= data_sz;
>> >> >> >+
>> >> >> >+                       if (rte_pktmbuf_chain(test_data->decomp_bufs[i],
>> >> >> >+                                       next_seg) < 0) {
>> >> >> >+                               RTE_LOG(ERR, USER1, "Could not chain mbufs\n");
>> >> >> >+                               return -1;
>> >> >> >+                       }
>> >> >> >+                       segs_per_mbuf++;
>> >> >> >+               }
>> >> >> >+
>> >> >> >+               /* Allocate data in output mbuf */
>> >> >> >+               test_data->comp_bufs[i] =
>> >> >> >+                       rte_pktmbuf_alloc(test_data->comp_buf_pool);
>> >> >> >+               if (test_data->comp_bufs[i] == NULL) {
>> >> >> >+                       RTE_LOG(ERR, USER1, "Could not allocate mbuf\n");
>> >> >> >+                       return -1;
>> >> >> >+               }
>> >> >> >+               data_addr = (uint8_t *) rte_pktmbuf_append(
>> >> >> >+                                       test_data->comp_bufs[i],
>> >> >> >+                                       test_data->seg_sz);
>> >> >> >+               if (data_addr == NULL) {
>> >> >> >+                       RTE_LOG(ERR, USER1, "Could not append data\n");
>> >> >> >+                       return -1;
>> >> >> >+               }
>> >> >> >+
>> >> >> >+               /* Chain mbufs if needed for output mbufs */
>> >> >> >+               for (j = 1; j < segs_per_mbuf; j++) {
>> >> >> >+                       struct rte_mbuf *next_seg =
>> >> >> >+
>> >> >> >+ rte_pktmbuf_alloc(test_data->comp_buf_pool);
>> >> >> >+
>> >> >> >+                       if (next_seg == NULL) {
>> >> >> >+                               RTE_LOG(ERR, USER1,
>> >> >> >+                                       "Could not allocate mbuf\n");
>> >> >> >+                               return -1;
>> >> >> >+                       }
>> >> >> >+
>> >> >> >+                       data_addr = (uint8_t *)rte_pktmbuf_append(next_seg,
>> >> >> >+                               test_data->seg_sz);
>> >> >> >+
>> >> >> >+                       if (data_addr == NULL) {
>> >> >> >+                               RTE_LOG(ERR, USER1, "Could not append data\n");
>> >> >> >+                               return -1;
>> >> >> >+                       }
>> >> >> >+
>> >> >> >+                       if (rte_pktmbuf_chain(test_data->comp_bufs[i],
>> >> >> >+                                       next_seg) < 0) {
>> >> >> >+                               RTE_LOG(ERR, USER1, "Could not chain mbufs\n");
>> >> >> >+                               return -1;
>> >> >> >+                       }
>> >> >> >+               }
>> >> >> >+       }
>> >> >> >+
>> >> >> >+       return 0;
>> >> >> >+}
>> >> >> >+
>> >> >> >+static void
>> >> >> >+free_bufs(struct comp_test_data *test_data) {
>> >> >> >+       uint32_t i;
>> >> >> >+
>> >> >> >+       for (i = 0; i < test_data->total_bufs; i++) {
>> >> >> >+               rte_pktmbuf_free(test_data->comp_bufs[i]);
>> >> >> >+               rte_pktmbuf_free(test_data->decomp_bufs[i]);
>> >> >> >+       }
>> >> >> >+       rte_free(test_data->comp_bufs);
>> >> >> >+       rte_free(test_data->decomp_bufs); }
>> >> >> >+
>> >> >> >+static int
>> >> >> >+main_loop(struct comp_test_data *test_data, uint8_t level,
>> >> >> >+                       enum rte_comp_xform_type type,
>> >> >> >+                       uint8_t *output_data_ptr,
>> >> >> >+                       size_t *output_data_sz,
>> >> >> >+                       unsigned int benchmarking) {
>> >> >> >+       uint8_t dev_id = test_data->cdev_id;
>> >> >> >+       uint32_t i, iter, num_iter;
>> >> >> >+       struct rte_comp_op **ops, **deq_ops;
>> >> >> >+       void *priv_xform = NULL;
>> >> >> >+       struct rte_comp_xform xform;
>> >> >> >+       size_t output_size = 0;
>> >> >> >+       struct rte_mbuf **input_bufs, **output_bufs;
>> >> >> >+       int res = 0;
>> >> >> >+       int allocated = 0;
>> >> >> >+
>> >> >> >+       if (test_data == NULL || !test_data->burst_sz) {
>> >> >> >+               RTE_LOG(ERR, USER1,
>> >> >> >+                       "Unknow burst size\n");
>> >> >> >+               return -1;
>> >> >> >+       }
>> >> >> >+
>> >> >> >+       ops = rte_zmalloc_socket(NULL,
>> >> >> >+               2 * test_data->total_bufs * sizeof(struct rte_comp_op *),
>> >> >> >+               0, rte_socket_id());
>> >> >> >+
>> >> >> >+       if (ops == NULL) {
>> >> >> >+               RTE_LOG(ERR, USER1,
>> >> >> >+                       "Can't allocate memory for ops strucures\n");
>> >> >> >+               return -1;
>> >> >> >+       }
>> >> >> >+
>> >> >> >+       deq_ops = &ops[test_data->total_bufs];
>> >> >> >+
>> >> >> >+       if (type == RTE_COMP_COMPRESS) {
>> >> >> >+               xform = (struct rte_comp_xform) {
>> >> >> >+                       .type = RTE_COMP_COMPRESS,
>> >> >> >+                       .compress = {
>> >> >> >+                               .algo = RTE_COMP_ALGO_DEFLATE,
>> >> >> >+                               .deflate.huffman = test_data->huffman_enc,
>> >> >> >+                               .level = level,
>> >> >> >+                               .window_size = test_data->window_sz,
>> >> >> >+                               .chksum = RTE_COMP_CHECKSUM_NONE,
>> >> >> >+                               .hash_algo = RTE_COMP_HASH_ALGO_NONE
>> >> >> >+                       }
>> >> >> >+               };
>> >> >> >+               input_bufs = test_data->decomp_bufs;
>> >> >> >+               output_bufs = test_data->comp_bufs;
>> >> >> >+       } else {
>> >> >> >+               xform = (struct rte_comp_xform) {
>> >> >> >+                       .type = RTE_COMP_DECOMPRESS,
>> >> >> >+                       .decompress = {
>> >> >> >+                               .algo = RTE_COMP_ALGO_DEFLATE,
>> >> >> >+                               .chksum = RTE_COMP_CHECKSUM_NONE,
>> >> >> >+                               .window_size = test_data->window_sz,
>> >> >> >+                               .hash_algo = RTE_COMP_HASH_ALGO_NONE
>> >> >> >+                       }
>> >> >> >+               };
>> >> >> >+               input_bufs = test_data->comp_bufs;
>> >> >> >+               output_bufs = test_data->decomp_bufs;
>> >> >> >+       }
>> >> >> >+
>> >> >> >+       /* Create private xform */
>> >> >> >+       if (rte_compressdev_private_xform_create(dev_id, &xform,
>> >> >> >+                       &priv_xform) < 0) {
>> >> >> >+               RTE_LOG(ERR, USER1, "Private xform could not be
>> created\n");
>> >> >> >+               res = -1;
>> >> >> >+               goto end;
>> >> >> >+       }
>> >> >> >+
>> >> >> >+       uint64_t tsc_start, tsc_end, tsc_duration;
>> >> >> >+
>> >> >> >+       tsc_start = tsc_end = tsc_duration = 0;
>> >> >> >+       if (benchmarking) {
>> >> >> >+               tsc_start = rte_rdtsc();
>> >> >> >+               num_iter = test_data->num_iter;
>> >> >> >+       } else
>> >> >> >+               num_iter = 1;
>> >> >> Looks like in same code we're doing benchmarking and functional
>> >> validation.
>> >> >> It can be reorganised to keep validation test separately like done
>> >> >> in crypto_perf.
>> >> >
>> >> >TJ: Ok, makes sense. However in the interests of getting this into
>> >> >the
>> >> >18.11 release I'd like to defer this refactoring and the remainder
>> >> >of your
>> >> comments below to the next release.
>> >> >
>> >> >
>> >> >Next comments - WIP
>> >> >
>> >> >
>> >> >Br, Tomek