Performance: Window Service

[bw-solana]

It takes ~25us per shred to go through the insertion routine and ~90us per shred to go through everything (recovery, flushing metas, etc.).

But as we pump more shreds through the system, these numbers start getting worse. For example, when looking at instances where we pump >45k shreds through, the per shred insertion cost approaches 100us and overall loops cost 300us.

At 100us per shred, we’re only getting 4k shreds through per slot, which means 2k data shreds, which means maybe 8k small transactions per block.

This whole system needs to be made much more efficient.

We also should not be relying so heavily on blockstore as a communication mechanism with repair and replay. We also need to limit how long we can spend in a loop before signaling (we currently flush all writes at the end but can spend 1 second or more in this loop in extreme cases).