feat: simple bench target for measuring DB reads using ReadView

[netrome]

Linked Issues/PRs

• closes chore: Benchmark multi-get implementation #2422

Description

This PR adds a benchmark target for measuring end to end query times for transaction queries. The benchmark is a simple binary that submits a bunch of transactions, and later queries these transactions by owner several times. It is intended to be used together with benchmarking tools such as hyperfine.

[netrome]

Interesting observation, the current benchmark is slower with the multi-get implementation than with the plain db lookups.

On my build server I get the following results (having saved the end_to_end_query_times_reference and end_to_end_query_times_multi_get binaries from previous runs with cargo bench --bench end_to_end_query_times)

hyperfine /tmp/end_to_end_query_times_reference /tmp/end_to_end_query_times_multi_get
Benchmark 1: /tmp/end_to_end_query_times_reference
  Time (mean ± σ):      7.052 s ±  0.200 s    [User: 7.726 s, System: 0.937 s]
  Range (min … max):    6.768 s …  7.246 s    10 runs

Benchmark 2: /tmp/end_to_end_query_times_multi_get
  Time (mean ± σ):      7.172 s ±  0.119 s    [User: 7.833 s, System: 0.990 s]
  Range (min … max):    6.901 s …  7.288 s    10 runs

Summary
  '/tmp/end_to_end_query_times_reference' ran
    1.02 ± 0.03 times faster than '/tmp/end_to_end_query_times_multi_get'

[netrome]

Bumped the number of blocks (and therefore number of transactions to query) by a factor 10, but I still can't observe a notable difference between the multi-get implementation and the previous one.

hyperfine /tmp/end_to_end_query_times_long_reference /tmp/end_to_end_query_times_long_multi_get
Benchmark 1: /tmp/end_to_end_query_times_long_reference
  Time (mean ± σ):     68.403 s ±  0.345 s    [User: 77.285 s, System: 8.956 s]
  Range (min … max):   67.802 s … 69.064 s    10 runs

Benchmark 2: /tmp/end_to_end_query_times_long_multi_get
  Time (mean ± σ):     68.726 s ±  0.499 s    [User: 77.652 s, System: 8.834 s]
  Range (min … max):   67.866 s … 69.568 s    10 runs

Summary
  '/tmp/end_to_end_query_times_long_reference' ran
    1.00 ± 0.01 times faster than '/tmp/end_to_end_query_times_long_multi_get'

This could be due to a number of factors.

1. One loose theory is that most (or all) of the data is still within the RocksDB memtable, and that this makes the speed gain by batching with multi-get negligible.
2. Another theory is that the overhead of boxing the iterators outweighs the performance gain of the multi-get operation for these workloads.
3. It could also just be the case that database lookup times isn't the hot path of the current workloads. If so, we could consider writing more isolated benchmarks (though we already have the db_lookup_times), but we should also question the value of this optimization if it's not in the hot path for the graphql queries since that's where it's used right now.
4. Anything else not considered yet.

[netrome]

I've rewritten the workload right now to only measure the database interactions through the ReadDatabase type, which isolates the effect of the multi-get implementation more than the full GraphQL interaction. This just confirms what I've already observed in the end-to-end benchmarks; the multi-get implementation is slower (for this workload).

(With the implementation from 4e7d3b357cf3b7edec95ea2859df63cf14459a35)

hyperfine /tmp/batch_lookup_times_coins_reference /tmp/batch_lookup_times_coins_multi_get
Benchmark 1: /tmp/batch_lookup_times_coins_reference
  Time (mean ± σ):      6.418 s ±  0.131 s    [User: 6.258 s, System: 0.059 s]
  Range (min … max):    6.281 s …  6.647 s    10 runs

Benchmark 2: /tmp/batch_lookup_times_coins_multi_get
  Time (mean ± σ):      7.365 s ±  0.082 s    [User: 6.163 s, System: 1.105 s]
  Range (min … max):    7.239 s …  7.539 s    10 runs

[netrome]

Running the latest benches on my laptop only amplifies the difference:

Baseline:

hyperfine target/release/deps/end_to_end_query_times-731a676331ca9c3d
Benchmark 1: target/release/deps/end_to_end_query_times-731a676331ca9c3d
  Time (mean ± σ):     13.720 s ±  0.228 s    [User: 13.642 s, System: 0.059 s]
  Range (min … max):   13.427 s … 14.121 s    10 runs

With multi-get

hyperfine target/release/deps/read_view_get_coins-ff72e5522e7d0f70
Benchmark 1: target/release/deps/read_view_get_coins-ff72e5522e7d0f70
  Time (mean ± σ):     24.500 s ±  1.737 s    [User: 23.881 s, System: 0.562 s]
  Range (min … max):   22.666 s … 27.756 s    10 runs

I've also experimented with disabling the rocksdb block cache, which didn't make any big difference for this workload.

For this workload, I can now safely say that the multi-get implementation is strictly worse. I suspect that this is due to the overhead caused by using dynamic dispatch for the iterators, but have yet to confirm this theory.

[netrome]

Reading up on multi-get, one performance benefit is reduced lock contention for concurrent reads - so I'll adjust the workload to see if I can measure anything interesting with many concurrent reads. https://github.com/facebook/rocksdb/wiki/MultiGet-Performance

[netrome]

Reading up on multi-get, one performance benefit is reduced lock contention for concurrent reads - so I'll adjust the workload to see if I can measure anything interesting with many concurrent reads. https://github.com/facebook/rocksdb/wiki/MultiGet-Performance

Multi-get is still slower. Measured at commit 5c261ee9897282ab4e5bbee6f79a030bac5d217d.

Reference:

hyperfine target/release/deps/read_view_get_coins-ff72e5522e7d0f70
Benchmark 1: target/release/deps/read_view_get_coins-ff72e5522e7d0f70
  Time (mean ± σ):      1.638 s ±  0.120 s    [User: 29.559 s, System: 0.233 s]
  Range (min … max):    1.389 s …  1.743 s    10 runs

Multi-get:

hyperfine target/release/deps/read_view_get_coins-ff72e5522e7d0f70
Benchmark 1: target/release/deps/read_view_get_coins-ff72e5522e7d0f70
  Time (mean ± σ):      1.962 s ±  0.039 s    [User: 20.858 s, System: 3.210 s]
  Range (min … max):    1.887 s …  2.019 s    10 runs

[xgreenx]

What batch size are you using by default? Based on the benchmarks from the @rymnc in #2142 , the multi-get is slower for small batch size. Have you tried batch size 250/500/1000?

[xgreenx]

Also, maybe you want to try to select random coins, not sequential.