Network Load Balancing: Emerging Techniques and Innovative Insights

Classic Load‑Balancing Schemes

CONGA – flow‑level granularity; carries link‑load information in packets; updates are propagated hop‑by‑hop.

Hula – flow‑level granularity; maps each flow to the next hop with highest path utilization; uses periodic probing.

DRILL – packet‑level granularity; leverages queue depth and per‑port load; computes path choice directly in the switch.

Hermes – flow‑level granularity; relies on RTT and ECN signals; retrieves information directly from the network.

MP‑RDMA: Enabling Multi‑Path Transport for RDMA

Background : RoCEv2‑based single‑path RDMA suffers from severe throughput loss (≈90 % drop) when packet loss reaches 0.5 % and from low bandwidth utilization due to ECMP hash collisions.

Motivation : NIC memory is limited, preventing storage of per‑path state and causing cache misses.

Multi‑path ACK clock – a single global congestion window is driven by ACKs; no per‑path state is kept, allowing dynamic traffic distribution across paths.

Out‑of‑order‑aware path selection – a 64‑bit bitmap tracks reordered packets; experiments show a 99.9 % reduction in out‑of‑order packets and only a 3.94 % throughput loss when a link is slowed.

Synchronous operation guarantee – a sync flag enforces memory‑update ordering; with 80 % of operations marked synchronous, throughput exceeds single‑path RDMA by 16.28 %.

Results : Under link failures, MP‑RDMA achieves 2–4× higher throughput and improves overall network utilization by 47 % (Lu et al., 2018)[5].

ConWeave: In‑Network Reordering Support for RDMA

Background : End‑host reordering adds latency in heavily loaded datacenters.

Motivation : Programmable switches (e.g., Intel Tofino) can perform reordering inside the network.

Cautious reroute – rerouting occurs only when (a) the current path is congested, (b) an uncongested alternative exists, and (c) packets from the previous reroute have already arrived at the destination ToR.

RTT‑based real‑time congestion monitoring.

Low‑overhead path selection – uses in‑band signals instead of active probing.

In‑network reordering – destination ToR marks packets with TAIL or REROUTED flags to restore order.

Results : Controlling packet order during path changes improves reliability and throughput (Song et al., 2023)[6].

Proteus: Multi‑Level Signals for Lossless Datacenter Networks

Background : PFC pauses become frequent with complex traffic, breaking traditional load‑balancing signals.

Motivation : Accurate path state is needed during PFC pauses.

Multi‑level signaling – combines RTT‑level signals (RTT, link utilization) with sub‑RTT signals (cumulative stall time, CST) to classify paths as non‑congested, congested, or indeterminate due to PFC.

Dynamic PFC‑pause handling – short pauses are tolerated; for long pauses, packets are rerouted based on CST using a “late is better than never” policy.

Results : Proteus maintains stable throughput and mitigates head‑of‑line blocking under PFC (Hu et al., 2024)[7].

CAVER: Hunting Less‑Congested Paths for RDMA

Background : RDMA’s strict ordering makes TCP‑based load balancing ineffective; out‑of‑order packets drastically reduce throughput.

Motivation : Precise, scalable congestion awareness is required.

ACK‑based congestion table – ACK packets carry per‑hop congestion metrics; switches update a Path Congestion Table similarly to CONGA’s DRE.

Vector protocol – ensures every switch can compute the optimal path from the latest congestion data.

Fast convergence – path selection converges within one RTT, minimizing oscillation.

Results : ACK‑driven updates markedly improve the accuracy and efficiency of RDMA load balancing (Deng et al., 2024)[8].