Noise Can Help: Accurate and Efficient Per-flow Latency Measurement without Packet Probing and Time Stamping Dept. of Computer Science and Engineering Michigan State University East Lansing, Michigan, 48824, USA Muhammad Shahzad Alex X. Liu
2 Latency Matters! Applications ─ Financial trading ─ HPC Architecture ─ CDNs ─ Data centers
“When considering how to reduce latency, the first step is to measure it.” (Joanne Kinsella, Head of Portfolio, British Telecom)
4 Flow 1 = 1ms Flow 2 = 1ms Flow 3 = 1.25ms Flow 4 = 4ms Aggregate Latency Measurement ─ Measure average latency ─ Guaranteeing average ≠ Guaranteeing each Per-flow Latency Measurement ─ Measure latency of each flow Applications ─ ISP operators ─ ISP customers Types of Latency Measurements Aggregate Latency = 1.7ms
5 Prior Art and Limitations Aggregate Latency Measurement ─ LDA [SIGCOMM’09] ─ FineComb [SIGMETRICS’11] Per-flow Latency Measurement ─ RLI [SIGCOMM’10]: active probes ─ MAPLE [IMC’12]: timestamps Commercial Solution ─ Corvil’s latency monitoring devices ─ USD 180,000 for a 2 × 10Gbps box
6 Problem Statement
7 Basic Idea
8 Recording Phase: a naïve solution One counter per flow: 1-1 mapping Problem ─ Overflow vs. Underutilization Reason ─ 1-1 mapping: flows counters
9 Recording Phase: COLATE Cost per packet: ─ One hash computation ─ One memory update
10 Querying Phase Latency of Flow ? Take average
11 Optimal Parameter Selection
12 Optimal Parameter Selection
13 Performance Evaluation Network Traces Simulated queue traversal to get departure timestamps ─ RED queue management strategy TraceDurationNo. of Packets No. of Flows CHIC5 mins37.3M3.01M ICSI41.1 hrs46.9M0.387M DC1.08 hrs19.9M0.439M
14 Accuracy
15 Comparison with RLI Implemented RLI (SIGCOMM’10)
16 Advantages over Prior-Art Proposed an accurate and efficient per-flow latency measurement scheme ─ Reliable ─ Passive ─ Scalable ─ Efficient: Memory and Computations ─ Flexible More in the paper ─ Standard deviation in latencies of packets in a flow ─ Theoretical development
17 Questions?