On 17 December 2015 at 00:50, Mike Holmes <mike.hol...@linaro.org> wrote:
> with apply and build
>
> Using patch: lng-odp_PATCH_v2_test_performance_add_crypto_test.mbox
> Trying to apply patch
> Patch applied
>
> WARNING: 'Imediately' may be misspelled - perhaps 'Immediately'?
> #645: FILE: test/performance/odp_crypto.c:582:
> + print_mem("Imediately encrypted packet", mem,
>
> total: 0 errors, 1 warnings, 0 checks, 958 lines checked
>
> And compiling
>
> odp_crypto.c: In function ‘create_session_from_config’:
> odp_crypto.c:428:33: error: storage size of ‘ses_create_rc’ isn’t known
> enum odp_crypto_ses_create_err ses_create_rc;
> ^
> odp_crypto.c:428:33: error: unused variable ‘ses_create_rc’
> [-Werror=unused-variable]
> odp_crypto.c: In function ‘main’:
> odp_crypto.c:777:17: error: implicit declaration of function
> ‘odp_cpumask_def_worker’ [-Werror=implicit-function-declaration]
> num_workers = odp_cpumask_def_worker(&cpumask,
> ^
> odp_crypto.c:777:3: error: nested extern declaration of
> ‘odp_cpumask_def_worker’ [-Werror=nested-externs]
> num_workers = odp_cpumask_def_worker(&cpumask,
> ^
> odp_crypto.c:795:4: error: too few arguments to function
> ‘odph_linux_pthread_create’
> odph_linux_pthread_create(&thr, &cpumask,
> ^
> In file included from odp_crypto.c:20:0:
> ../../helper/include/odp/helper/linux.h:67:5: note: declared here
> int odph_linux_pthread_create(odph_linux_pthread_t *thread_tbl,
> ^
> cc1: all warnings being treated as errors
>
> So it looks like it needs a rebase.
>
[Alex] Patch is a bit old, I'll rebase to latest ODP master.
>
> On 4 November 2015 at 10:07, Nicolas Morey-Chaisemartin <nmo...@kalray.eu>
> wrote:
>
>>
>> On 11/04/2015 09:12 AM, alexandru.badici...@linaro.org wrote:
>> > From: Alexandru Badicioiu <alexandru.badici...@linaro.org>
>> >
>> > Test program to measure crypto operation performance.
>> > Measures the time required to launch and get the result
>> > of ODP crypto operations in async and sync API mode with
>> > configurable payload size, crypto algorithms and iteration
>> > number.
>> > Both asynchronous scheduled and polled are supported.
>> > In scheduled mode a separate worker thread is used to
>> > get the crypto completion events.
>> > Output packet can be reused as the input of the next
>> > operation or a new packet can be allocated for each
>> > operation.
>> > Based on a previous work :
>> > https://lists.linaro.org/pipermail/lng-odp/2014-September/003251.html.
>> >
>> > Signed-off-by: Alexandru Badicioiu <alexandru.badici...@linaro.org>
>> > v1
>> > ---
>> > - replaced cspeed with crypto
>> > - removed unused parameter core_count
>> > v2
>> > --
>> > - alphabetical order in Makefile.am
>> > - replaced pool params memset with init call
>> >
>> > ---
>> > test/performance/Makefile.am | 5 +-
>> > test/performance/odp_crypto.c | 935
>> ++++++++++++++++++++++++++++++++++++++++++
>> > 2 files changed, 939 insertions(+), 1 deletion(-)
>> > create mode 100644 test/performance/odp_crypto.c
>>
>> > +/**
>> > + * Snap current time values and put them into 'rec'.
>> > + */
>> > +static void
>> > +fill_time_record(time_record_t *rec)
>> > +{
>> > + gettimeofday(&rec->tv, NULL);
>> > + getrusage(RUSAGE_SELF, &rec->ru_self);
>> > + getrusage(RUSAGE_THREAD, &rec->ru_thread);
>> This is *very* linux oriented and not platform agnostic.
>> getrusage is POSIX 2001 (though we don't support it) but RUSAGE_THREAD is
>> _GNU_SOURCE and kernel > 2.6.26
>>
>> Couldn't we use ODP API to keep track of time ?
>>
>
> Nicolas, in the call today we put out the idea that it would be better to
> use an odp time api, but that maybe it was ok to use linux for now.
> Re reading the patch I think this is problem since it will break on your
> platform and not just be less clean, is that correct ?
>
>
>> > +int main(int argc, char *argv[])
>> > +{
>> > + crypto_args_t cargs;
>> > + odp_pool_t pool;
>> > + odp_queue_param_t qparam;
>> > + odp_pool_param_t params;
>> > + odph_linux_pthread_t thr;
>> > + odp_queue_t out_queue = ODP_QUEUE_INVALID;
>> > + thr_arg_t thr_arg;
>> > + int num_workers = 1;
>> > + odp_cpumask_t cpumask;
>> > + char cpumaskstr[ODP_CPUMASK_STR_SIZE];
>> > +
>> > + /* Parse and store the application arguments */
>> > + parse_args(argc, argv, &cargs);
>> > +
>> > + /* Init ODP before calling anything else */
>> > + if (odp_init_global(NULL, NULL)) {
>> > + app_err("ODP global init failed.\n");
>> > + exit(EXIT_FAILURE);
>> > + }
>> > +
>> > + /* Init this thread */
>> > + odp_init_local(ODP_THREAD_WORKER);
>> > +
>> > + /* Create packet pool */
>> > + odp_pool_param_init(¶ms);
>> > + params.pkt.seg_len = SHM_PKT_POOL_BUF_SIZE;
>> > + params.pkt.len = SHM_PKT_POOL_BUF_SIZE;
>> > + params.pkt.num = SHM_PKT_POOL_SIZE / SHM_PKT_POOL_BUF_SIZE;
>> > + params.type = ODP_POOL_PACKET;
>> > + pool = odp_pool_create("packet_pool", ¶ms);
>> > +
>> > + if (pool == ODP_POOL_INVALID) {
>> > + app_err("packet pool create failed.\n");
>> > + exit(EXIT_FAILURE);
>> > + }
>> > + odp_pool_print(pool);
>> > +
>> > + if (cargs.schedule) {
>> > + qparam.sched.prio = ODP_SCHED_PRIO_DEFAULT;
>> > + qparam.sched.sync = ODP_SCHED_SYNC_NONE;
>> > + qparam.sched.group = ODP_SCHED_GROUP_ALL;
>> > + out_queue = odp_queue_create("crypto-out",
>> > + ODP_QUEUE_TYPE_SCHED,
>> &qparam);
>> > + } else if (cargs.poll) {
>> > + out_queue = odp_queue_create("crypto-out",
>> > + ODP_QUEUE_TYPE_POLL, NULL);
>> > + }
>> > + if (cargs.schedule || cargs.poll) {
>> > + if (out_queue == ODP_QUEUE_INVALID) {
>> > + app_err("crypto-out queue create failed.\n");
>> > + exit(EXIT_FAILURE);
>> > + }
>> > + }
>> > +
>> > + if (cargs.schedule) {
>> > + printf("Run in async scheduled mode\n");
>> > +
>> > + thr_arg.crypto_args = cargs;
>> > + thr_arg.crypto_alg_config = cargs.alg_config;
>> > + num_workers = odp_cpumask_def_worker(&cpumask,
>> > + num_workers);
>> > + (void)odp_cpumask_to_str(&cpumask, cpumaskstr,
>> > + sizeof(cpumaskstr));
>> > + printf("num worker threads: %i\n",
>> > + num_workers);
>> > + printf("first CPU: %i\n",
>> > + odp_cpumask_first(&cpumask));
>> > + printf("cpu mask: %s\n",
>> > + cpumaskstr);
>> It would be good to have an option for the number of workerds (or a core
>> bitmask), as was recently added to odp_l2fwd
>>
>
> Sounds sensible
>
>
>> > + } else if (cargs.poll) {
>> > + printf("Run in async poll mode\n");
>> > + } else {
>> > + printf("Run in sync mode\n");
>> > + }
>> > +
>> > + if (cargs.alg_config) {
>> > + if (cargs.schedule) {
>> > + odph_linux_pthread_create(&thr, &cpumask,
>> > + run_thr_func, &thr_arg);
>> > + odph_linux_pthread_join(&thr, num_workers);
>> > + } else {
>> > + run_measure_one_config(&cargs, cargs.alg_config);
>> > + }
>> Why is there only multithreading in schedule mode?
>>
>
> Not sure, Alex is there a technical reason ?
>
[Alex] Not sure I understand the question - I guess the intent was to ask
"why is there only one thread". Or the question was "why there's a separate
thread to run the test"?
>
>
>> _______________________________________________
>> lng-odp mailing list
>> lng-odp@lists.linaro.org
>> https://lists.linaro.org/mailman/listinfo/lng-odp
>>
>
>
>
> --
> Mike Holmes
> Technical Manager - Linaro Networking Group
> Linaro.org <http://www.linaro.org/> *│ *Open source software for ARM SoCs
>
>
>
_______________________________________________
lng-odp mailing list
lng-odp@lists.linaro.org
https://lists.linaro.org/mailman/listinfo/lng-odp
