hyperledger-labs / fabric-smart-client / 26634201560

71%

Build Type

push

github

Committed by web-flow

Commit Message

perf(bft): per-orderer channel-slot pool to fix semaphore mutex starvation (#1490)

BFTBroadcaster.getConnection serialized every caller through the single
sync.Mutex inside a shared golang.org/x/sync/semaphore.Weighted, and spun
on a 1s-timeout Acquire retry loop when the pool was empty. Under high
broadcast concurrency this collapses into mutex starvation: every
broadcaster contends on the one semaphore regardless of which orderer it
dials, and a caller whose context is cancelled mid-acquire busy-spins
because the loop has no ctx.Done() exit.

Replace the global semaphore with per-orderer state:
  - a fixed-capacity channel of slot tokens that bounds connections per
    orderer without a mutex shared across orderers, and
  - a buffered pool of idle *Connection for reuse.

getConnection becomes a three-stage, non-spinning acquire (try pool, try
slot, then block on a select over pool / slot / ctx.Done()). SendAndRecv
wraps the send+recv round-trip in a derived context so a hung Recv is
cancelled and torn down instead of leaking.

Tests: adapt the partial-failure discard test to the per-orderer slot
accounting, and add a regression test that getConnection returns promptly
on a cancelled context.

Observed in our deployment (c=1000 stress): submit p99 6.18s -> 1.59s,
678 TPS sustained.

Signed-off-by: Evan <evanyan@sign.global>

Coverage Stats

84 of 118 new or added lines in 2 files covered. (71.19%)

11 existing lines in 4 files now uncovered.

28425 of 39904 relevant lines covered (71.23%)

2515.62 hits per line