# ignacio (jsign) - Update 10

In my previous update, I've created a benchmark that compares face to face inserting 1k, 5k and 10k key-values in an in-memory Merkle Patricia Trie and Verkle Trie. The strategy is to start moving to more end to end benchmarks to keep finding optimization opportunities.

In this case, I made changes in the three layers:
- `go-ethereum`: https://github.com/gballet/go-ethereum/pull/147
- `go-verkle`: https://github.com/gballet/go-verkle/pull/314
- `go-ipa`: https://github.com/crate-crypto/go-ipa/pull/31

In `go-ethereum` all three combined changes had the following impact in a 16 core machine:

name                               old time/op    new time/op    delta
TriesRandom/VKT/1000_accounts-16     63.1ms ± 4%    20.8ms ± 3%  -67.10%  (p=0.000 n=19+10)
TriesRandom/VKT/5000_accounts-16      279ms ± 4%      67ms ± 1%  -75.83%  (p=0.000 n=19+9)
TriesRandom/VKT/10000_accounts-16     562ms ± 2%     128ms ± 1%  -77.15%  (p=0.000 n=20+9)

name                               old alloc/op   new alloc/op   delta
TriesRandom/VKT/1000_accounts-16     9.49MB ± 0%    3.57MB ± 0%  -62.39%  (p=0.000 n=20+10)
TriesRandom/VKT/5000_accounts-16     43.5MB ± 0%    13.0MB ± 0%  -70.15%  (p=0.000 n=20+10)
TriesRandom/VKT/10000_accounts-16    87.4MB ± 0%    26.3MB ± 0%  -69.93%  (p=0.000 n=20+10)

name                               old allocs/op  new allocs/op  delta
TriesRandom/VKT/1000_accounts-16      32.8k ± 0%     35.5k ± 0%   +8.05%  (p=0.000 n=20+10)
TriesRandom/VKT/5000_accounts-16       151k ± 0%      144k ± 0%   -4.90%  (p=0.000 n=20+10)
TriesRandom/VKT/10000_accounts-16      303k ± 0%      285k ± 0%   -5.98%  (p=0.000 n=20+10)

In summary, around ~75% performance improvement (4.4x speedup) in the refernece benchmarks.
The performance improvement scales with the number of cores in the machine. This was ran less powerful setups with [good results too](https://gist.github.com/jsign/e064a7a96e762f6e4729a666b4b86bec).

These changes also had promising results in paths that we didn't focus optimizing but since they use the same underlying code paths, we got other gains "for free".

For example, [in the `ProofCalculation` benchmark in `go-verkle`](https://github.com/gballet/go-verkle/pull/314):

name                                        old time/op    new time/op    delta
ProofCalculation-16                            1.87s ± 2%     0.08s ± 3%  -95.75%  (p=0.003 n=5+7)

None of the PRs are meregd yet. There were many changes that already had a pass of review from Guillaume. We've decided to [wait for other PR that is baking in `go-verkle`](https://github.com/gballet/go-verkle/pull/295) since that can affect some changes that I did in this PR, so we want to see after how we reorganize this work.

Guillaume will be OOO for some weeks, so we'll freeze a bit this work direction until he's back.

I already have plans on working on a different direction resuming some chat I had with Dankrad some weeks ago, so this I'm not blocked.