Preview: Operating Systems
summary
Shows AVX vector code (256/512-bit AVX) forcing cores into lower clock states and proposes scheduler detection plus compensation for hyperthread victims. Presents a hybrid event queue to cut Linux packet latency versus one-to-one and one-to-N queues. Introduces a huge page allocator targeting 2 MB huge pages to reduce TLB pressure while limiting fragmentation. Proposes a NUMA-aware kernel to avoid remote DRAM churn. Implements Homa in-kernel for fair comparison with TCP. Designs enclaves beyond 256 MB and hardware virtualization for computational SSDs and a network stack beating RDMA’s 1–2 µs tail latencies.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:34A. Process scheduling
- 1:25Background: AVX instructions
- 1:59Problem: unfairness from AVX instructions
- 3:06Paper #2
- 3:32Problem: latency in current event queues
- 4:52B. Memory management
- 5:52Background: translation and hugepages
- 7:18Problem: fragmentation from hugepages
- 8:26Background: non-uniform memory access
- 9:08Problem: NUMA slows monolithic kernels
- 10:23Background: multi-tiered memory (MTM)
- 11:14Problem: inefficient MTM usage
- 12:15C. Network management
- 13:06Paper #1
- 13:22Background: Homa transport protocol
- 14:27Problem: Homa is not tested practically
- 15:05D: Novel hardware designs
- 15:40Background: enclaves
- 16:37Problem: existing enclaves do not scale
- 17:30Background: computational storage
- 18:13Problem: CSD virtualization is inefficient
- 18:50Paper #3
- 19:00Background: remote procedure calls
- 19:47Problem: existing network stacks are slow
- 20:32Conclusion
related talks
20mKeynote: MDS, Fallout, Zombieland & Linux
01h 11mPerformance engineering - being friendly to your hardware
24mPREEMPT_RT Q&A with tglx
11mAdvanced Parallel Memory Virtualization
45mDynamic vCPU priority boosting in KVM for latency sensitive workloads
37mMeta’s CXL Journey and Learnings in Linux
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints
16mXkernel: Principled Performance Tunability of Operating System Kernels