Add some performance optimizations #944

Jake-Shadle · 2024-05-07T15:26:07Z

This adds parallel options to improve deserialization speed of Endpoint, and Resource.

Endpoint now has a from_proto and into_proto methods so that the current From impls continue to work as they do currently. This differences in these methods are that the protobufs now actually differentiate between ipv4/ipv6 and host, so we no longer need to parse a string in all cases, and instead just use a u32 for ipv4 and a pair of u64s for ipv6. This also saves on serialization though I'm not measuring that.
Resource::Listener => Resource::FilterChain. This resource doesn't see as much traffic as Cluster, but it was still low hanging fruit. The envoy listener protobuf is massive overkill since we use literally 1 field from it, in addition to filter having more fields than we used, and also not enough.

Lines 824 to 825 in e44ca38

    
           // TODO: keep the label across xDS 
        
           label: None,

Cluster can now take advantage of the Endpoint optimizations to give pretty significant wins.

misc               fastest       │ slowest       │ median        │ mean          │ samples │ iters
├─ endpoint                      │               │               │               │         │
│  ├─ faster                     │               │               │               │         │
│  │  ├─ Ip        229.7 ns      │ 18.82 µs      │ 230.7 ns      │ 259.5 ns      │ 1000    │ 1000
│  │  │            1.775 GB/s    │ 17.84 MB/s    │ 1.56 GB/s     │ 1.475 GB/s    │         │
│  │  ╰─ Name      281.4 ns      │ 995.3 ns      │ 290.2 ns      │ 294.5 ns      │ 1000    │ 8000
│  │               2.216 GB/s    │ 618.8 MB/s    │ 2.149 GB/s    │ 2.111 GB/s    │         │
│  ╰─ slower                     │               │               │               │         │
│     ├─ Ip        289.7 ns      │ 6.812 µs      │ 310.7 ns      │ 322 ns        │ 1000    │ 1000
│     │            1.297 GB/s    │ 49.31 MB/s    │ 1.261 GB/s    │ 1.158 GB/s    │         │
│     ╰─ Name      351.6 ns      │ 1.102 µs      │ 364.1 ns      │ 366.5 ns      │ 1000    │ 8000
│                  1.729 GB/s    │ 551.2 MB/s    │ 1.669 GB/s    │ 1.653 GB/s    │         │
╰─ resource                      │               │               │               │         │
   ├─ faster                     │               │               │               │         │
   │  ├─ Cluster   2.518 ms      │ 4.482 ms      │ 3.01 ms       │ 3.019 ms      │ 1000    │ 1000
   │  │            799.1 MB/s    │ 449 MB/s      │ 668.6 MB/s    │ 666.6 MB/s    │         │
   │  ╰─ Listener  22.95 µs      │ 285.5 µs      │ 23.41 µs      │ 23.92 µs      │ 1000    │ 1000
   │               95.15 MB/s    │ 7.648 MB/s    │ 93.27 MB/s    │ 91.29 MB/s    │         │
   ╰─ slower                     │               │               │               │         │
      ├─ Cluster   3.475 ms      │ 5.588 ms      │ 3.589 ms      │ 3.602 ms      │ 1000    │ 1000
      │            586.3 MB/s    │ 364.5 MB/s    │ 567.6 MB/s    │ 565.5 MB/s    │         │
      ╰─ Listener  23.81 µs      │ 232.7 µs      │ 24.59 µs      │ 25.26 µs      │ 1000    │ 1000
                   90.02 MB/s    │ 9.212 MB/s    │ 87.16 MB/s    │ 84.87 MB/s    │         │

As noted, this doesn't actually change anything since these are backwards incompatible changes.

github-actions · 2024-05-07T15:26:24Z