nft_set_pipapo: Introduce AVX2-based lookup implementation
If the AVX2 set is available, we can exploit the repetitive characteristic of this algorithm to provide a fast, vectorised version by using 256-bit wide AVX2 operations for bucket loads and bitwise intersections. In most cases, this implementation consistently outperforms rbtree set instances despite the fact they are configured to use a given, single, ranged data type out of the ones used for performance measurements by the nft_concat_range.sh kselftest. That script, injecting packets directly on the ingoing device path with pktgen, reports, averaged over five runs on a single AMD Epyc 7402 thread (3.35GHz, 768 KiB L1D$, 12 MiB L2$), the figures below. CONFIG_RETPOLINE was not set here. Note that this is not a fair comparison over hash and rbtree set types: non-ranged entries (used to have a reference for hash types) would be matched faster than this, and matching on a single field only (which is the case for rbtree) is also significantly faster. However, it's not possible at the moment to choose this set type for non-ranged entries, and the current implementation also needs a few minor adjustments in order to match on less than two fields. ---------------.-----------------------------------.------------. AMD Epyc 7402 | baselines, Mpps | this patch | 1 thread |___________________________________|____________| 3.35GHz | | | | | | 768KiB L1D$ | netdev | hash | rbtree | | | ---------------| hook | no | single | | pipapo | type entries | drop | ranges | field | pipapo | AVX2 | ---------------|--------|--------|--------|--------|------------| net,port | | | | | | 1000 | 19.0 | 10.4 | 3.8 | 4.0 | 7.5 +87% | ---------------|--------|--------|--------|--------|------------| port,net | | | | | | 100 | 18.8 | 10.3 | 5.8 | 6.3 | 8.1 +29% | ---------------|--------|--------|--------|--------|------------| net6,port | | | | | | 1000 | 16.4 | 7.6 | 1.8 | 2.1 | 4.8 +128% | ---------------|--------|--------|--------|--------|------------| port,proto | | | | | | 30000 | 19.6 | 11.6 | 3.9 | 0.5 | 2.6 +420% | ---------------|--------|--------|--------|--------|------------| net6,port,mac | | | | | | 10 | 16.5 | 5.4 | 4.3 | 3.4 | 4.7 +38% | ---------------|--------|--------|--------|--------|------------| net6,port,mac, | | | | | | proto 1000 | 16.5 | 5.7 | 1.9 | 1.4 | 3.6 +26% | ---------------|--------|--------|--------|--------|------------| net,mac | | | | | | 1000 | 19.0 | 8.4 | 3.9 | 2.5 | 6.4 +156% | ---------------'--------'--------'--------'--------'------------' A similar strategy could be easily reused to implement specialised versions for other SIMD sets, and I plan to post at least a NEON version at a later time. Signed-off-by: Stefano Brivio <sbrivio@redhat.com> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Showing
This diff is collapsed.
Please register or sign in to comment