Aggregating Funnels

Aggregating Funnels (Roh, Wei, Ruppert et al., PPoPP 2025) is a software-combining layer that replaces a hardware FAA hotspot with batched FAA operations across threads. Plugged into LCRQ, it improves throughput by up to 2.5× at high thread counts, making LCRQ + Aggregating Funnels the fastest published MPMC queue as of 2025.

The problem it solves

FAA beats CAS decisively under contention — but FAA itself is not free. A LOCK XADD on a single contended cache line still serializes through the cache coherence protocol, and at 100+ threads even FAA becomes a bottleneck. The hardware physically cannot grant atomic increments faster than one cache-coherence round-trip.

Aggregating Funnels addresses this by combining FAA calls in software before they hit the hardware. A funnel is a tree of nodes; threads contend at the leaves, partial increments are merged on the way up, and a single FAA is performed at the root for the combined increment. Each thread then retrieves its individual slot index from the combined result.

Why it works

The trick is that FAA is associative: incrementing by 1 a hundred times produces the same shared-index result as incrementing by 100 once. A funnel exploits this to amortize the cache-coherence cost of the contended atomic across many participants. Latency for an individual operation goes up slightly, but throughput goes up dramatically — and at the thread counts where the funnel is profitable, the per-thread latency reduction from avoided contention exceeds the funnel overhead anyway.

This is the same combining-tree idea that goes back to Yew, Tzeng, and Lawrie's NYU Ultracomputer work (1987), but tuned for modern cache hierarchies rather than network-on-chip routing.

Performance

On the standard MPMC queue benchmarks at 144 threads on a 4×18-core Intel Xeon:

• Up to 2.5× higher throughput than vanilla LCRQ
• Improvement scales with thread count — minimal at 16 threads, peak at 100+
• Negligible overhead at low thread counts where contention is already low

This makes LCRQ + Aggregating Funnels the new ceiling for strict-FIFO MPMC, supplanting plain LCRQ for the first time in 12 years.

Where it goes from here

Aggregating Funnels is a layer, not a queue — the same combining technique applies to any contended FAA. The PPoPP 2025 paper demonstrates it on counters and stacks as well. The natural next step is plugging it into LPRQ for portable architectures, and into wCQ to give a wait-free queue with the same scalability ceiling.

See also

• lcrq — the queue this layer was originally designed for
• the-fastest-queue — full hierarchy
• faa-vs-cas — why FAA was the right primitive to combine