Re: [PATCH v3 1/2] bpf: add a longest prefix match trie map implementation

On Wed, Jan 18, 2017 at 03:30:14PM +0100, David Herrmann wrote:

Hi

On Sat, Jan 14, 2017 at 5:55 PM, Alexei Starovoitov [off-list ref] wrote:

quoted

Another alternative is to extend samples/bpf/map_perf_test
It has perf tests for most map types today (including lru)
and trie would be natural addition there.
I would prefer this latter option.

I hooked into gettid() and installed a simple kprobe bpf program that
searches for an entry in an lpm trie. The bpf program does either 0,
1, 8, or 32 lookups in a row (always the same element). The trie has
size 0, 1, or 8192. The data is below. The results vary by roughly 5%
on every run.

A single gettid() syscall with an empty bpf program takes roughly
6.5us on my system. Lookups in empty tries take ~1.8us on first try,
~0.9us on retries. Lookups in tries with 8192 entries take ~7.1us (on
the first _and_ any subsequent try).

thanks. please add them to commit log.

https://gist.github.com/dvdhrm/4c90e61a1c39746d5c55ab9e0e29315e

looks good. please add it as an extra patch to this set.
and add
#pragma clang loop unroll(full)
before
+       for (i = 0; i < 8; ++i)
+               bpf_map_lookup_elem(&lpm_trie_map_alloc, &key);
to make sure it's unrolled regardless of the version of llvm.

Trie-size: 0
#Lookups: 0
    0:lpm_perf kmalloc 9,230,321 events per sec
    -> 6.5us / syscall

Trie-size: 1
#Lookups: 1
    0:lpm_perf kmalloc 7,224,508 events per sec
    -> 8.3us / syscall

Trie-size: 1
#Lookups: 8
    0:lpm_perf kmalloc 4,152,740 events per sec
    -> 14.4us / syscall

Trie-size: 1
#Lookups: 32
    0:lpm_perf kmalloc 1,713,415 events per sec
    -> 35.0us / syscall

Trie-size: 8192
#Lookups: 1
    0:lpm_perf kmalloc 4,369,138 events per sec
    -> 13.7us / syscall

Trie-size: 8192
#Lookups: 8
    0:lpm_perf kmalloc 943,849 events per sec
    -> 63.6us / syscall

Trie-size: 8192
#Lookups: 32
    0:lpm_perf kmalloc 271,737 events per sec
    -> 220.8us / syscall

so it's the same lookup done 32-times per each syscall.
Since the latency of single lookup in 8 vs 32 experiments
are pretty close, it means that the 32 lookup numbers
amortize the cost of syscall pretty well and no need to go higher.
So I would use 32 as a loop count in stress_lpm_trie_map_alloc()

Thanks!