Exploring Linux Memory Usage and IO Performance in Cloud Native Applications
summary
Uses fio to measure buffered vs O_DIRECT IO on an EC2 c5.large (burst 20k IOPS for 30 min, sustained 4k IOPS, 125 MB/s). Emphasizes the kernel’s MemAvailable metric and demonstrates concrete runs: cached 2 GB reads hit ~4.5 GB/s with ~0.4 ms latency, uncached 2–4 GB reads fall to hundreds of seconds when active set > available memory. Explains writes are throttled by dirty limits (~15% of available memory), not symmetric with reads, and shows how memory allocation causes the hockey‑stick collapse in IO performance.
AI-written summary · about
chapters (5)
related talks
38mScaling IO-bound application in the cloud
25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte Platform
26mLet's demystify memory
01h 36mLinux Memory Management and Containers
48mPersistent Memory Usage within Linux Environment
40mBenchmarking: Repeatable & Comparable
40mInspecting and Optimizing Memory Usage in Linux
40mAchieving the Ultimate Performance with KVM