Fast Software Cache Design for Network Appliances
summary
Replaces Open vSwitch’s microflow table with a new table layout and eviction policy: Bounded Linear Probing lets keys spill into the next bucket to boost occupancy, trading off only 1.5 consecutive cache-line reads on average instead of two distant reads; Probabilistic Bubble LRU bubbles hot entries forward with infrequent promotions to avoid extra metadata. Paper includes analysis, microbenchmarks and an OvS prototype that yields up to 15% throughput improvement (and much larger gains on skewed workloads).
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:26Flow Caching in Open vSwitch
- 3:33Basic Cache Design
- 6:07Increase Set-Associativity
- 7:18More Candidate Buckets
- 9:01Our Solution: Bounded Linear Probing (BLP)
- 11:19Why BLP is Better Than Set-Assoc.?
- 12:51Qualitative Comparison
- 13:55Better Cache Replacement
- 15:06PBLRU: Bubbling
- 18:17Evaluation
- 19:32Lookup Latency and Hit Rate
- 20:51Throughput (Skewed) Skewed (0-0.99)
- 21:58Summary
related talks
14mScaling Open vSwitch with a Computational Cache
17mPushing the Limits of In-Network Caching for Key-Value Stores
17mCo-Designing Traffic Control with NVMe-oF for Disaggregated Storage: A Comparative Study
18mSeer: Enabling Future-Aware Online Caching in Networked Systems
14mPolicyCache: Intra-flow Learning in Congestion Control
16mONCache: A Cache-Based Low-Overhead Container Overlay Network
25mResQ: Enabling SLOs in Network Function Virtualization
28mLoom: Flexible and Efficient NIC Packet Scheduling