Multi-address client: optimize client selection

[idelpivnitskiy]

Motivation:

selectClient attracted my attention after I saw it can take up to 10% on the flame graph.

Modifications:

• Change CachingKeyFactory to UrlKeyFactory;
• Remove an extra layer of caching for UrlKeys;
• Eagerly compute UrlKey.hashCode because it's used >1 time;
• Remove allocation of HostAndPort from the hot path because it's required only to create a new client once;

Result:

Client selection reduced from 63.788 ± 1.268 ns/op to 14.512 ± 0.233 ns/op.

---

I tried a few different approaches. All of them are shown here as independent commits.

Benchmark                                                    (lowercase)  Mode  Cnt   Score   Error  Units
Original                                                            true  avgt    5  63.788 ± 1.268  ns/op
Original                                                           false  avgt    5  64.920 ± 0.517  ns/op
v1: UrlKey caches hashCode                                          true  avgt    5  44.918 ± 0.685  ns/op
v1: UrlKey caches hashCode                                         false  avgt    5  45.352 ± 0.510  ns/op
v2: remove urlKeyCache, key as a string                             true  avgt    5  81.459 ± 1.087  ns/op
v2: remove urlKeyCache, key as a string                            false  avgt    5  82.457 ± 1.015  ns/op
v3: remove urlKeyCache, key as UrlKey                               true  avgt    5  19.767 ± 0.350  ns/op
v3: remove urlKeyCache, key as UrlKey                              false  avgt    5  19.813 ± 0.048  ns/op
v4: remove urlKeyCache, key as UrlKey, don't allocate HostAndPort   true  avgt    5  14.512 ± 0.233  ns/op
v4: remove urlKeyCache, key as UrlKey, don't allocate HostAndPort  false  avgt    5  14.808 ± 0.324  ns/op

[bryce-anderson]

I like the simplification, that alone is worth it for me.

I am curious about the benchmarks: are they already committed somewhere or are they local? I'm specifically curious whether a lot of the benefit comes from the port being the first thing checked if the ports are set as random numbers vs the most common 80/443.

[idelpivnitskiy]

I am curious about the benchmarks: are they already committed somewhere or are they local?

See commits in this PR. I added the benchmark and removed it later because it required some internal changes for multi-address client builder.

I'm specifically curious whether a lot of the benefit comes from the port being the first thing checked if the ports are set as random numbers vs the most common 80/443.

I used a fixed port bcz int check seems trivial to make any difference.

[bryce-anderson]

Looks good after you address the spotbugs complaints regarding the omission of the type and null checking in UrlKey.equals(..). (I don't care if they're ignored or added back.)

[idelpivnitskiy]

Verified with the benchmark that it doesn't make any difference, added them back