jlokier / paprika

An experimental storage for Nethermind, removing the whole Trie abstraction and acting as a Trie and a database at once

🌶️ Paprika

Paprika provides a custom implementation of the Patricia tree used in Ethereum. It aims at delivering the following:

1. no external KV storage
2. no managed memory overhead - Paprika uses almost no managed memory as only Span<byte>, stackallocated variables and uses no managed representation of the tree
3. atomic commit
4. the history of N-th last versions of the db

No external KV storage means that Paprika is self-sufficient in regards to writing and reading data from disk and requires no third-party databases like RocksDB or others.

No managed memory overhead means that Paprika uses almost no managed memory. Only Span<byte> and stackallocated variables are used. There's no representation of the tree stored in the managed memory. Also, it uses [module: SkipLocalsInit] to not clean any memory before accessing it.

Atomic commit means atomic in the meaning of ACID. If there's a failure during the commit the database shall not be corrupted and shall represent the previous state.

History is preserved using additional metadata pages, that keep the root and the pages that were marked as abandoned.

Design

Visit docs for further information.

Benchmarks

After the redesign to be page-oriented, the benchmarks use bigger numbers

1. 160 million pairs, written in a single transaction, which shall simulate the initial population/sync

In each case, the key is 32 bytes long. The value is 32 bytes long as well.

General considerations

Memory Profiling

Almost no managed memory is used

160 million pairs

Writing 160 million 32bytes -> 32bytes mappings

Wrote 10,000,000 items, DB usage is at 5.01% which gives 1.00GB out of allocated 20GB
Wrote 20,000,000 items, DB usage is at 9.85% which gives 1.97GB out of allocated 20GB
Wrote 30,000,000 items, DB usage is at 11.41% which gives 2.28GB out of allocated 20GB
Wrote 40,000,000 items, DB usage is at 15.05% which gives 3.01GB out of allocated 20GB
Wrote 50,000,000 items, DB usage is at 18.85% which gives 3.77GB out of allocated 20GB
Wrote 60,000,000 items, DB usage is at 21.10% which gives 4.22GB out of allocated 20GB
Wrote 70,000,000 items, DB usage is at 24.86% which gives 4.97GB out of allocated 20GB
Wrote 80,000,000 items, DB usage is at 28.02% which gives 5.60GB out of allocated 20GB
Wrote 90,000,000 items, DB usage is at 30.84% which gives 6.17GB out of allocated 20GB
Wrote 100,000,000 items, DB usage is at 34.42% which gives 6.88GB out of allocated 20GB
Wrote 110,000,000 items, DB usage is at 37.41% which gives 7.48GB out of allocated 20GB
Wrote 120,000,000 items, DB usage is at 40.52% which gives 8.10GB out of allocated 20GB
Wrote 130,000,000 items, DB usage is at 43.89% which gives 8.78GB out of allocated 20GB
Wrote 140,000,000 items, DB usage is at 46.91% which gives 9.38GB out of allocated 20GB
Wrote 150,000,000 items, DB usage is at 50.14% which gives 10.03GB out of allocated 20GB
Writing of 160,000,000.00 items took 00:08:19.8204173 giving a throughput 320,114.00 items/s

Committing to disk took 00:00:10.4085680

Reading of 160,000,000.00 items took 00:08:24.8362414 giving a throughput 316,934.00 items/s