Source-linked AI summary
QPS-ToR: A Parallel Iterative Switching Algorithm for Reconfigurable Optical Datacenter Switching
Dongzhao Song, Qianru Yu, Jun Xu
TL;DR
Growing DCN traffic calls for switching that delivers high throughput and delay, while traffic-oblivious forwarding and NegotiaToR’s single-iteration scheduler have throughput limits. QPS-ToR adapts sliding-window SW-QPS into a two-round Request-Grant pipeline; simulations report up to 36% higher throughput and 82% lower FCT than NegotiaToR, and consistent gains over both baselines above 0.5 load.
Problem
Traffic-oblivious VLB limits throughput to around 60%, while NegotiaToR’s single-iteration iSLIP achieves below 60% without fabric speedup [19].
Source: §1.1 · p. 1, §1.2 · p. 2Method
QPS-ToR replaces NegotiaToR’s single-iteration iSLIP engine with SW-QPS, using queue-proportional requests and grants over a sliding window.
Source: §2.2 · p. 3, §6 · p. 5Results
Up to 36% higher throughput and up to 82% lower FCT than NegotiaToR, plus up to 22% higher throughput and up to 50% lower FCT than RotorNet, are reported.
Source: §6 · pp. 5–6Takeaways & Limitations
QPS-ToR steers optical capacity toward heavily loaded ToR pairs, draining large backlogs before substantial queue buildup occurs.
Source: §5.2 · p. 5
Abstract
from arXiv · showhide
Reconfigurable optical data center networks (RODCNs) have emerged as a promising solution for scaling DCN capacity, yet their scheduling mechanisms remain a performance bottleneck: traffic-oblivious schemes inherently limit throughput, while the state-of-the-art traffic-aware scheme, NegotiaToR, uses single-iteration iSLIP as its scheduling engine, which limits throughput to around 60% and treats all source-destination pairs with equal priority regardless of queue length. We propose QPS-ToR, which replaces NegotiaToR's scheduling logic with SW-QPS, a sliding-window algorithm originally proposed for crossbar scheduling that achieves around 90% throughput with a single low-complexity iteration. QPS-ToR operates within NegotiaToR's existing workflow, requiring only a revision of the scheduling cycle from three-step Request-Grant-Accept (RGA) to two-step Request-Grant (RG) with a sliding window mechanism. In flow-level simulations with 128 ToRs under realistic datacenter workloads, QPS-ToR achieves up to 36% higher throughput and 82% lower flow completion time (FCT) compared to NegotiaToR, and consistently outperforms RotorNet, a representative traffic-oblivious scheme, on the parallel network topology.
1. INTRODUCTION
Growing DCN traffic calls for optical switching that can deliver high throughput and delay performance; existing traffic-oblivious and traffic-aware schedulers each leave performance gaps. QPS-ToR adapts SW-QPS for NegotiaToR and reports improved throughput and flow completion time.
1.1 Traffic-oblivious solutions
- Traffic-oblivious VLB-based forwarding limits throughput to around 60%, since most traffic takes two hops to reach its destination.RotorNet is identified as a representative traffic-oblivious scheme.Source: §1.1 · p. 1
1.2 Traffic-aware solutions
- NegotiaToR runs only one iSLIP RGA iteration across three consecutive epochs, with reported throughput below 60% absent fabric speedup [19].Multiple iterations cannot fit within an epoch because inter-ToR round trips can exceed an epoch.Source: §1.2 · p. 2
1.3 QPS-ToR Algorithm
- QPS-ToR adapts SW-QPS [31] and revises NegotiaToR’s scheduling workflow to address the read-after-write hazard at negligible performance cost.SW-QPS was proposed for input-queued crossbar scheduling.Source: §1.3 · p. 2
- Up to 36% higher throughput and up to 82% lower FCT than NegotiaToR are reported for QPS-ToR.The proposed scheduler replaces NegotiaToR’s single-iteration iSLIP engine with SW-QPS.Source: §1.3 · p. 2
2. BACKGROUND
The background describes NegotiaToR’s epoch pipeline and the SW-QPS mechanism that QPS-ToR adapts. The contrast is a three-round RGA pipeline versus queue-proportional request and grant decisions over a sliding window.
2.1 NegotiaToR workflow
- NegotiaToR’s predefined phase pipelines R, G, and A across epochs to produce one matching per epoch for the scheduled phase.The predefined phase acts as the control plane, while the scheduled phase carries bulk transfers using the completed matching.Source: §2.1 · p. 2
2.2 QPS and SW-QPS
- QPS sends one request per input with probability proportional to the corresponding VOQ length, and each output grants the request with the largest queue.Because each input sends only one request, QPS does not need a separate accept phase.Source: §2.2 · p. 3
- SW-QPS adds edges to T pending matchings in each request-grant iteration, granting the earliest slot when both ports are available.The oldest matching graduates for use next, while a new empty matching enters the sliding window.Source: §2.2 · p. 3
3. QPS-TOR ALGORITHM
QPS-ToR replaces NegotiaToR’s pipelined RGA operation with a pipelined RG operation. It addresses the resulting read-after-write hazard by partitioning the sliding-window epochs across two scheduler instances.
3. QPS-TOR ALGORITHM
- QPS-ToR’s RG pipeline has two segments, R and G, rather than NegotiaToR’s three RGA segments, reducing pipeline depth to two.The paper states this also improves latency beyond the benefit of SW-QPS as a switching algorithm.Source: §3 · p. 3
- Concurrent request and grant processing can double-book an input port when its request reports availability that a grant has already consumed.The hazard arises because requests report a snapshot of future availability while grants add edges using an earlier snapshot.Source: §3 · p. 3
- Two SW-QPS scheduler instances operate on disjoint epoch sets to prevent double booking in the sliding window.The partition nominally halves input availability, but experiments find negligible impact on throughput and latency.Source: §3 · p. 3
4. RELATED WORK
The related-work discussion contrasts optical-only RODCN scheduling with hybrid circuit and packet switching, which adds an electronic packet switch. It notes both hybrid scaling costs and the continuing need for good optical scheduling.
4. RELATED WORK
- Hybrid switching pairs an OCS with a lower-bandwidth EPS that avoids reconfiguration delay, whereas RODCN schedules one fixed-duration matching per epoch without a dedicated EPS helper.RODCN epochs last microseconds, compared with milliseconds for hybrid switching.Source: §4 · p. 4
- At N = 96, only QPS-Fit computes a hybrid OCS schedule within a 3 ms epoch; other algorithms take one to three orders longer.The paper attributes QPS-Fit’s speed to its massively parallelable design.Source: §4 · p. 4
- Negligible OCS guardbands moderate time spent on mice flows, but improved optical hardware alone does not ensure better FCT or throughput.The paper presents QPS-ToR as a scheduling algorithm for this remaining need.Source: §4 · p. 4
5. EVALUATION
In a 128-ToR evaluation across three datacenter workloads, QPS-ToR outperforms NegotiaToR and RotorNet on FCT and throughput. Its queue-aware scheduling steers optical capacity toward heavily loaded ToR pairs, with benefits that vary by workload.
5.2 Evaluation Result
- At loads above 0.5, QPS-ToR reduces average FCT by roughly 72% relative to NegotiaToR and 35% relative to RotorNet under Trace 1.At load 0.7, the reductions reach 82% and 50%, respectively; under Trace 2 at high loads, FCT improves by up to 68% over NegotiaToR and 37% over RotorNet.Source: §5.2 · p. 5, Figure 3 · p. 5
- Under lighter-tailed Trace 3, QPS-ToR improves FCT by roughly 50% over NegotiaToR and 21% over RotorNet, while its throughput advantage remains consistent across all loads.The reduced prevalence of elephant flows limits the benefit of queue-aware scheduling.Source: §5.2 · p. 5
5.1 Evaluation setup
- QPS-ToR steers optical capacity toward heavily loaded ToR pairs, draining large backlogs before substantial queue buildup, without workload-specific parameter tuning.The evaluation uses 128 ToRs, eight parallel 128×128 AWGR switches, and offered loads up to the full 800 Gbps ToR uplink capacity.Source: §5.1 · pp. 4–5, §5.2 · p. 5
5.2 Evaluation Result
- At loads above 0.5 under Trace 1, QPS-ToR improves throughput by approximately 24% over NegotiaToR and 8% over RotorNet.Under Trace 2, gains are 22% over NegotiaToR and 15% over RotorNet above load 0.5, reaching 36% and 22% at load 0.9.Source: §5.2 · p. 5, Figure 3 · p. 5
- Across all workloads above load 0.5, QPS-ToR outperforms both baselines on FCT, and it achieves higher throughput than both across workloads and load levels.Figure 3 compares FCT and throughput under the three workloads.Source: Figure 3 · p. 5, §5.2 · p. 5
6. CONCLUSION
QPS-ToR replaces NegotiaToR’s single-iteration iSLIP engine with SW-QPS and adapts its scheduling workflow. In simulations with 128 ToRs and three realistic workloads, it improves FCT and throughput over both NegotiaToR and RotorNet.
6. CONCLUSION
- In flow-level simulations with 128 ToRs under three realistic datacenter workloads, QPS-ToR reduces FCT by up to 82% versus NegotiaToR and 50% versus RotorNet, while improving throughput by up to 36% and 22%, respectively.It outperforms both baselines at loads above 0.5 across all traffic distributions.Source: §6 · pp. 5–6
- QPS-ToR replaces NegotiaToR’s single-iteration iSLIP engine with SW-QPS, a sliding-window queue-proportional sampling algorithm adapted from crossbar scheduling.Source: §6 · p. 5
- The scheduler retains NegotiaToR’s epoch workflow, reduces the pipeline from three-step RGA to two-step RG, and interleaves two independent instances on disjoint epoch subsets.Source: §6 · p. 5