[PATCH v4 3/7] eal: add lcore variable performance test

[Mattias Rönnblom]

On 2024-09-16 12:52, Mattias Rönnblom wrote:
> Add basic micro benchmark for lcore variables, in an attempt to assure
> that the overhead isn't significantly greater than alternative
> approaches, in scenarios where the benefits aren't expected to show up
> (i.e., when plenty of cache is available compared to the working set
> size of the per-lcore data).
> 

Here are some test results for a Raptor Cove @ 3,2 GHz (GCC 11):

  + ------------------------------------------------------- +
  + Test Suite : lcore variable perf autotest
  + ------------------------------------------------------- +
Latencies [TSC cycles/update]
Modules/Variables  Static array  Thread-local Storage  Lcore variables
                 1           3.9           5.5              3.7
                 2           3.8           5.5              3.8
                 4           4.9           5.5              3.7
                 8           3.8           5.5              3.8
                16          11.3           5.5              3.7
                32          20.9           5.5              3.7
                64          23.5           5.5              3.7
               128          23.2           5.5              3.7
               256          23.5           5.5              3.7
               512          24.1           5.5              3.7
              1024          25.3           5.5              3.9
  + TestCase [ 0] : test_lcore_var_access succeeded
  + ------------------------------------------------------- +


The reason for TLS being slower than lcore variables (which in turn 
relies on TLS for lcore id lookup) is the lazy initialization 
conditional that is imposed on variant. Could that be avoided (which is 
module-dependent I suppose), it beats lcore variables at ~3.0 cycles/update.

I must say I'm surprised to see lcore variables doing this good, at 
these very modest working set sizes. Probably, you can stay at near-zero 
L1 misses with lcore variables (and TLS), but start missing the L1 with 
static arrays.

> Signed-off-by: Mattias Rönnblom <mattias.ronnblom at ericsson.com>
> 
> --
> 
> PATCH v4:
>   * Rework the tests to be a little less unrealistic. Instead of a
>     single dummy module using a single variable, use a number of
>     variables/modules. In this way, differences in cache effects may
>     show up.
>   * Add RTE_CACHE_GUARD to better mimic that static array pattern.
>     (Morten Brørup)
>   * Show latencies as TSC cycles. (Morten Brørup)
> ---
>   app/test/meson.build           |   1 +
>   app/test/test_lcore_var_perf.c | 244 +++++++++++++++++++++++++++++++++
>   2 files changed, 245 insertions(+)
>   create mode 100644 app/test/test_lcore_var_perf.c
> 
> diff --git a/app/test/meson.build b/app/test/meson.build
> index 48279522f0..d4e0c59900 100644
> --- a/app/test/meson.build
> +++ b/app/test/meson.build
> @@ -104,6 +104,7 @@ source_file_deps = {
>       'test_kvargs.c': ['kvargs'],
>       'test_latencystats.c': ['ethdev', 'latencystats', 'metrics'] + sample_packet_forward_deps,
>       'test_lcore_var.c': [],
> +    'test_lcore_var_perf.c': [],
>       'test_lcores.c': [],
>       'test_link_bonding.c': ['ethdev', 'net_bond',
>           'net'] + packet_burst_generator_deps + virtual_pmd_deps,
> diff --git a/app/test/test_lcore_var_perf.c b/app/test/test_lcore_var_perf.c
> new file mode 100644
> index 0000000000..8b0abc771c
> --- /dev/null
> +++ b/app/test/test_lcore_var_perf.c
> @@ -0,0 +1,244 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2024 Ericsson AB
> + */
> +
> +#define MAX_MODS 1024
> +
> +#include <stdio.h>
> +
> +#include <rte_bitops.h>
> +#include <rte_cycles.h>
> +#include <rte_lcore_var.h>
> +#include <rte_per_lcore.h>
> +#include <rte_random.h>
> +
> +#include "test.h"
> +
> +struct mod_lcore_state {
> +	uint64_t a;
> +	uint64_t b;
> +	uint64_t sum;
> +};
> +
> +static void
> +mod_init(struct mod_lcore_state *state)
> +{
> +	state->a = rte_rand();
> +	state->b = rte_rand();
> +	state->sum = 0;
> +}
> +
> +static __rte_always_inline void
> +mod_update(volatile struct mod_lcore_state *state)
> +{
> +	state->sum += state->a * state->b;
> +}
> +
> +struct __rte_cache_aligned mod_lcore_state_aligned {
> +	struct mod_lcore_state mod_state;
> +
> +	RTE_CACHE_GUARD;
> +};
> +
> +static struct mod_lcore_state_aligned
> +sarray_lcore_state[MAX_MODS][RTE_MAX_LCORE];
> +
> +static void
> +sarray_init(void)
> +{
> +	unsigned int lcore_id = rte_lcore_id();
> +	int mod;
> +
> +	for (mod = 0; mod < MAX_MODS; mod++) {
> +		struct mod_lcore_state *mod_state =
> +			&sarray_lcore_state[mod][lcore_id].mod_state;
> +
> +		mod_init(mod_state);
> +	}
> +}
> +
> +static __rte_noinline void
> +sarray_update(unsigned int mod)
> +{
> +	unsigned int lcore_id = rte_lcore_id();
> +	struct mod_lcore_state *mod_state =
> +		&sarray_lcore_state[mod][lcore_id].mod_state;
> +
> +	mod_update(mod_state);
> +}
> +
> +struct mod_lcore_state_lazy {
> +	struct mod_lcore_state mod_state;
> +	bool initialized;
> +};
> +
> +/*
> + * Note: it's usually a bad idea have this much thread-local storage
> + * allocated in a real application, since it will incur a cost on
> + * thread creation and non-lcore thread memory usage.
> + */
> +static RTE_DEFINE_PER_LCORE(struct mod_lcore_state_lazy,
> +			    tls_lcore_state)[MAX_MODS];
> +
> +static inline void
> +tls_init(struct mod_lcore_state_lazy *state)
> +{
> +	mod_init(&state->mod_state);
> +
> +	state->initialized = true;
> +}
> +
> +static __rte_noinline void
> +tls_update(unsigned int mod)
> +{
> +	struct mod_lcore_state_lazy *state =
> +		&RTE_PER_LCORE(tls_lcore_state[mod]);
> +
> +	/* With thread-local storage, initialization must usually be lazy */
> +	if (!state->initialized)
> +		tls_init(state);
> +
> +	mod_update(&state->mod_state);
> +}
> +
> +RTE_LCORE_VAR_HANDLE(struct mod_lcore_state, lvar_lcore_state)[MAX_MODS];
> +
> +static void
> +lvar_init(void)
> +{
> +	unsigned int mod;
> +
> +	for (mod = 0; mod < MAX_MODS; mod++) {
> +		RTE_LCORE_VAR_ALLOC(lvar_lcore_state[mod]);
> +
> +		struct mod_lcore_state *state =
> +			RTE_LCORE_VAR_VALUE(lvar_lcore_state[mod]);
> +
> +		mod_init(state);
> +	}
> +}
> +
> +static __rte_noinline void
> +lvar_update(unsigned int mod)
> +{
> +	struct mod_lcore_state *state =
> +		RTE_LCORE_VAR_VALUE(lvar_lcore_state[mod]);
> +
> +	mod_update(state);
> +}
> +
> +static void
> +shuffle(unsigned int *elems, size_t len)
> +{
> +	size_t i;
> +
> +	for (i = len - 1; i > 0; i--) {
> +		unsigned int other = rte_rand_max(i + 1);
> +
> +		unsigned int tmp = elems[other];
> +		elems[other] = elems[i];
> +		elems[i] = tmp;
> +	}
> +}
> +
> +#define ITERATIONS UINT64_C(10000000)
> +
> +static inline double
> +benchmark_access(const unsigned int *mods, unsigned int num_mods,
> +		 void (*init_fun)(void), void (*update_fun)(unsigned int))
> +{
> +	unsigned int i;
> +	double start;
> +	double end;
> +	double latency;
> +	unsigned int num_mods_mask = num_mods - 1;
> +
> +	RTE_VERIFY(rte_is_power_of_2(num_mods));
> +
> +	if (init_fun != NULL)
> +		init_fun();
> +
> +	/* Warm up cache and make sure TLS variables are initialized */
> +	for (i = 0; i < num_mods; i++)
> +		update_fun(i);
> +
> +	start = rte_rdtsc();
> +
> +	for (i = 0; i < ITERATIONS; i++)
> +		update_fun(mods[i & num_mods_mask]);
> +
> +	end = rte_rdtsc();
> +
> +	latency = (end - start) / ITERATIONS;
> +
> +	return latency;
> +}
> +
> +static void
> +test_lcore_var_access_n(unsigned int num_mods)
> +{
> +	double sarray_latency;
> +	double tls_latency;
> +	double lvar_latency;
> +	unsigned int mods[num_mods];
> +	unsigned int i;
> +
> +	for (i = 0; i < num_mods; i++)
> +		mods[i] = i;
> +
> +	shuffle(mods, num_mods);
> +
> +	sarray_latency =
> +		benchmark_access(mods, num_mods, sarray_init, sarray_update);
> +
> +	tls_latency =
> +		benchmark_access(mods, num_mods, NULL, tls_update);
> +
> +	lvar_latency =
> +		benchmark_access(mods, num_mods, lvar_init, lvar_update);
> +
> +	printf("%17u %13.1f %13.1f %16.1f\n", num_mods, sarray_latency,
> +	       tls_latency, lvar_latency);
> +}
> +
> +/*
> + * The potential performance benefit of lcore variables compared to
> + * the use of statically sized, lcore id-indexed arrays are not
> + * shorter latencies in a scenario with low cache pressure, but rather
> + * fewer cache misses in a real-world scenario, with extensive cache
> + * usage. These tests are a crude simulation of such, using <N> dummy
> + * modules, each wiht a small, per-lcore state. Note however that
> + * these tests has very little non-lcore/thread local state, which is
> + * unrealistic.
> + */
> +
> +static int
> +test_lcore_var_access(void)
> +{
> +	unsigned int num_mods = 1;
> +
> +	printf("Latencies [TSC cycles/update]\n");
> +	printf("Modules/Variables  Static array  Thread-local Storage  "
> +	       "Lcore variables\n");
> +
> +	for (num_mods = 1; num_mods <= MAX_MODS; num_mods *= 2)
> +		test_lcore_var_access_n(num_mods);
> +
> +	return TEST_SUCCESS;
> +}
> +
> +static struct unit_test_suite lcore_var_testsuite = {
> +	.suite_name = "lcore variable perf autotest",
> +	.unit_test_cases = {
> +		TEST_CASE(test_lcore_var_access),
> +		TEST_CASES_END()
> +	},
> +};
> +
> +static int
> +test_lcore_var_perf(void)
> +{
> +	return unit_test_suite_runner(&lcore_var_testsuite);
> +}
> +
> +REGISTER_PERF_TEST(lcore_var_perf_autotest, test_lcore_var_perf);