Re: [PATCH v2 2/2] selftest/bpf/benchs: Add bpf_map benchmark

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Tue, May 24, 2022 at 12:53 AM Feng zhou <zhoufeng.zf@bytedance.com> wrote:
> +static void setup(void)
> +{
> +       struct bpf_link *link;
> +       int map_fd, i, max_entries;
> +
> +       setup_libbpf();
> +
> +       ctx.skel = bpf_map_bench__open_and_load();
> +       if (!ctx.skel) {
> +               fprintf(stderr, "failed to open skeleton\n");
> +               exit(1);
> +       }
> +
> +       link = bpf_program__attach(ctx.skel->progs.benchmark);
> +       if (!link) {
> +               fprintf(stderr, "failed to attach program!\n");
> +               exit(1);
> +       }
> +
> +       //fill hash_map
> +       map_fd = bpf_map__fd(ctx.skel->maps.hash_map_bench);
> +       max_entries = bpf_map__max_entries(ctx.skel->maps.hash_map_bench);
> +       for (i = 0; i < max_entries; i++)
> +               bpf_map_update_elem(map_fd, &i, &i, BPF_ANY);
> +}
...
 +SEC("fentry/" SYS_PREFIX "sys_getpgid")
> +int benchmark(void *ctx)
> +{
> +       u32 key = bpf_get_prandom_u32();
> +       u64 init_val = 1;
> +
> +       bpf_map_update_elem(&hash_map_bench, &key, &init_val, BPF_ANY);
> +       return 0;
> +}

This benchmark is artificial at its extreme.
First it populates the map till max_entries and then
constantly bounces off the max_entries limit in a bpf prog.
Sometimes random_u32 will be less than max_entries
and map_update_elem will hit a fast path,
but most of the time it will fail to alloc_htab_elem()
and will fail to map_update_elem.

It does demonstrate that percpu_free_list is inefficient
when it's empty, but there is no way such a microbenchmark
justifies optimizing this corner case.

If there is a production use case please code it up in
a benchmark.

Also there is a lot of other overhead: syscall and atomic-s.
To stress map_update_elem please use a for() loop inside bpf prog.