workload/tpcc: pre-compute random data for initial data load strings

[danhhz]

This is explictly allowed by the spec only for initial data load in
4.3.2.1:

For the purpose of populating the initial database only, random numbers
can be generated by selecting entries in sequence from a set of at least
10,000 pregenerated random numbers. This technique cannot be used for
the field O_OL_CNT.

name                             old time/op    new time/op    delta
InitialData/tpcc/warehouses=1-8     368ms ± 9%     229ms ± 1%  -37.77%  (p=0.000 n=15+15)

name                             old speed      new speed      delta
InitialData/tpcc/warehouses=1-8   299MB/s ± 9%   481MB/s ± 1%  +60.50%  (p=0.000 n=15+15)

name                             old alloc/op   new alloc/op   delta
InitialData/tpcc/warehouses=1-8     193kB ± 0%     126kB ± 0%  -34.80%  (p=0.000 n=14+14)

name                             old allocs/op  new allocs/op  delta
InitialData/tpcc/warehouses=1-8       592 ± 0%       458 ± 0%  -22.59%  (p=0.000 n=15+15)

Release note: None

[cockroach-teamcity]

This change is 

[tbg]

I'm missing some context here, but it from the review it seems that the resulting code is kind of vague about what's precomputed and what isn't. It'd be nice if that were somehow more obvious.

Reviewed 9 of 9 files at r1.
Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @danhhz)

---

pkg/workload/bench_test.go, line 54 at r1 (raw file):

		// PrecomputedRand. This makes the benchmark results an upper bound on
		// performance/allocations. TPCC, for example, special cases the default
		// seed to do some initialization only once which means it will be faster

Without the storing of the precomputation for seed=0, what would be the real-world effects? Wouldn't we precompute this only once per node per import? Trying to understand whether it's worth the cognitive complexity to make one seed special.

---

pkg/workload/tpcc/random.go, line 89 at r1 (raw file):

// randAString generates a random alphanumeric string of length between min and
// max inclusive. See 4.3.2.2.
func randAString(rng *tpccRand, a *bufalloc.ByteAllocator, min, max int) []byte {

Should these methods make it clear (in their name) that they're only allowed for initial data?

---

pkg/workload/workloadimpl/precomputedrand_test.go, line 81 at r1 (raw file):

	var randOffset int

	b.Run(`len=2`, func(b *testing.B) {

Nit: fmt.Sprintf("len=%d", shortLen) at which point why not unify the three subbenchmarks and range []int{shortLen, mediumLen, longLen} { ... }

[danhhz]

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @tbg)

---

pkg/workload/bench_test.go, line 54 at r1 (raw file):

Previously, tbg (Tobias Grieger) wrote…

Without the storing of the precomputation for seed=0, what would be the real-world effects? Wouldn't we precompute this only once per node per import? Trying to understand whether it's worth the cognitive complexity to make one seed special.

good point. tpcc generates a 10KiB string. i was thinking ahead to tpch which has a 300MiB one, but init once may not even be how we want to do that so I'll punt this till we hit it

---

pkg/workload/tpcc/random.go, line 89 at r1 (raw file):

Previously, tbg (Tobias Grieger) wrote…

Should these methods make it clear (in their name) that they're only allowed for initial data?

Done

---

pkg/workload/workloadimpl/precomputedrand_test.go, line 81 at r1 (raw file):

Previously, tbg (Tobias Grieger) wrote…

Nit: fmt.Sprintf("len=%d", shortLen) at which point why not unify the three subbenchmarks and range []int{shortLen, mediumLen, longLen} { ... }

Done

[tbg]

Reviewed 6 of 6 files at r2.
Reviewable status: complete! 0 of 0 LGTMs obtained

[danhhz]

Putting the offsets in the sync.Pool objects wasn't deterministic (and the test caught it, yay). Should be deterministic again. It's a tiny bit slower than the non-deterministic version but still much faster than what's in master. For some reason I also seem to have saved some allocs

[danhhz]

Thanks for the review!

bors r=tbg

[craig]

Build succeeded

• GitHub CI (Cockroach)