You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Ensure that transaction tokens are unique in the mempool benchmarks
Data of type `Token` is what a mempool `TestBlock` transaction (`Tx`) consumes
and produces. Previously, this token was a newtype around a `Word8`. There can
only be around 256 unque `Word8` tokens, which means that its likely for this
value to wrap around. In the current benchmark scenario, where we add linearly
dependent transactions with only one input and output to a mempool, this wrap
around was occurring and therefore the list of transaction that was being
generated had many duplicates. The type of tokens is therefore changed to be a
newtype around an `Int`, which won't wrap around (in a realistic benchmark).
Abovementioned change exposed subtle performance effects on the mempool. Most of
this can be attributed to the `Tx` type doubling as a `GenTxId`. The mempool
keeps track of these identifiers in a set. Since there were only around 256
unique transactions, this set was rather small. Changing tokens to be `Int`s
ensured that this set could grow much larger before becoming full, and that by
itself has subtle effects on the performance of the mempool.
Apart from changing the type of tokens, the mempool benchmark setup is reworked:
* Transactions that are used as benchmark inputs are pre-generated and fully
evaluated so that the generation work is not measured in the benchmarked
function itself.
* Instead of opening a mempool once and removing transactions after each
benchmark run, a new mempool is opened in each benchmark run. Removing
transactions proved to be prohibitively costly compared to the cost of opening
a new mempool.
* A new benchmark is added that measures just the mempool setup time. To obtain
the cost of just adding the transactions without cost of opening the mempool,
the time of this new benchmark can be manually subtracted from the full
benchmark time.
0 commit comments