Is Your Kernel Being Honest? Understanding & measuring low level bottlenecks
summary
A short demo program allocates 1 GB but pmap shows 1 GB virtual maps to 4 KB RSS, and touching that memory triggers ~250k page faults so that populating the pages costs about a third of a second of kernel time while the user loop only takes ~60 ms. Tools used are pmap, perf (perf stat/trace), simple alloc and instructions-per-time binaries and eBPF, and perf reveals perf stat shows 250k page faults costing ~345 ms kernel time.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 1:06Speaker introduction
- 2:10Talk agenda overview
- 2:32Defining IO in computing
- 3:59Linux memory fundamentals
- 4:42Memory allocation demo
- 6:09Using pmap tool
- 7:55Virtual vs physical memory
- 8:57Page faults explained
- 10:41Perf tool demonstration
- 11:48Page faults cost analysis
- 13:58Memory recap and impact
- 14:43CPU performance basics
- 15:26Measuring CPU utilization
- 17:31CPU efficiency metrics
- 18:14Instructions per time tool
- 20:42CPU pipelining explained
- 21:47IPC performance insights
- 22:51CPU recap and limitations
- 23:12What to measure in systems
- 24:56Performance monitoring tools
- 25:38eBPF for advanced tracing
- 26:20Capitalist tuning guide
- 28:09Closing gift and resources
- 29:11Q&A session
related talks
41mAre you out of memory, or have plenty to spare?
46mbpftrace internals
26mLet's demystify memory
31mFantastic tracepoints and where to find them
30m"Logs Told Us It Was Kernel – It Wasn't"
34mPrintk Should Be Your Last Resort
41mLies Programmers Believe about Memory
52mEffectively Measure and Reduce Kernel Latencies for Real-time Constraints