It is crucial that VersionedCrdsValue::insert_timestamp does not go
backward in time:
https://github.com/solana-labs/solana/blob/ec37a843a/core/src/crds.rs#L67-L79
Otherwise methods such as get_votes and get_epoch_slots_since will
break, which will break their downstream flow, including vote-listener
and optimistic confirmation:
https://github.com/solana-labs/solana/blob/ec37a843a/core/src/cluster_info.rs#L1197-L1215https://github.com/solana-labs/solana/blob/ec37a843a/core/src/cluster_info.rs#L1274-L1298
For that, Crds::new_versioned is intended to be called "atomically" with
Crds::insert_verioned (as the comment already says so):
https://github.com/solana-labs/solana/blob/ec37a843a/core/src/crds.rs#L126-L129
However, currently this is violated in the code. For example,
filter_pull_responses creates VersionedCrdsValues (with the current
timestamp), then acquires an exclusive lock on gossip, then
process_pull_responses writes those values to the crds table:
https://github.com/solana-labs/solana/blob/ec37a843a/core/src/cluster_info.rs#L2375-L2392
Depending on the workload and lock contention, the insert_timestamps may
well be in the past when these values finally are inserted into gossip.
To avoid such scenarios, this commit:
* removes Crds::new_versioned and Crd::insert_versioned.
* makes VersionedCrdsValue constructor private, only invoked in
Crds::insert, so that insert_timestamp is populated right before
insert.
This will improve insert_timestamp monotonicity as long as Crds::insert
is not called with a stalled timestamp. Following commits may further
improve this by calling timestamp() inside Crds::insert, and/or
switching to std::time::Instant which guarantees monotonicity.
filter_crds_values checks every crds filter against every hash value:
https://github.com/solana-labs/solana/blob/ee646aa7/core/src/crds_gossip_pull.rs#L432
which can be inefficient if the filter's bit-mask only matches small
portion of the entire crds table.
This commit shards crds values into separate tables based on shard_bits
first bits of their hash prefix. Given a (mask, mask_bits) filter,
filtering crds can be done by inspecting only relevant shards.
If CrdsFilter.mask_bits <= shard_bits, then precisely only the crds
values which match (mask, mask_bits) bit pattern are traversed.
If CrdsFilter.mask_bits > shard_bits, then approximately only
1/2^shard_bits of crds values are inspected.
Benchmarking on a gce cluster of 20 nodes, I see ~10% improvement in
generate_pull_responses metric, but with larger clusters, crds table and
2^mask_bits are both larger, so the impact should be more significant.