Round LRU time to nearest 32 seconds #4317
Conversation
|
Related issue with more details, for reference: #4169 |
| @@ -35,7 +35,10 @@ void nano::scheduler::bucket::pop () | |||
|
|
|||
| void nano::scheduler::bucket::push (uint64_t time, std::shared_ptr<nano::block> block) | |||
| { | |||
| queue.insert ({ time, block }); | |||
| const uint64_t bitmask = ~0b11111; | |||
| uint64_t rounded_time = time & bitmask; | |||
There was a problem hiding this comment.
For readability I would strongly suggest to keep a normal division. Our bottleneck is IO which is many orders of magnitude slower anyway, this division is noop when it comes to performance
|
This PR has been tested on the beta network. |
|
The performance improvement that was shown on the beta network turned out to be caused by a problem with one of the nodes that was fixed just after the LRU rounding version was installed. |
By rounding LRU time we are more likely to get AEC aligned across the network because blocks that are published very fast are grouped in 32 seconds "slots" that are sorted by hash.
This pr uses a 5 bit mask for rounding for best performance.
60 seconds rounding has been suggested but I chose 32 seconds to err on the side of caution and because Piotr mentioned that down to 15 seconds might be enough