Depends on how you split it up: if you're synchronizing 100 differences at a time it's on the order of 100,000 differences per second per core which is plenty f

Depends on how you split it up: if you're synchronizing 100 differences at a time it's on the order of 100,000 differences per second per core which is plenty fast for many applications. If you're working on thousands of differences at a time, then IBLT's communications efficiency may be a acceptable, and it is considerably faster. [You can always decrease the amount you feed into reconciliation at once by using hashing to subdivide your inputs, at a small loss to efficiency due to imbalancing.]