Performance optimization on Linux by Peter Toth

Performance optimizaton on LinuxBy Peter Toth

[email protected]

Intro

What do you want to learn today?

What level are you interested in? Any specific requests? Have you ever optimized Linux or an application for Linux?

Assuming everyone is familiar with Linux concepts and basic architecture

Agenda

MethodologyKnow your…Optimisable hardware components

Methodology

DocumentStart on a clean systemEmpty caches - warmup cachesAutomate testing (clear caches every time)Step-by-stepKeep those changes which are significantEstablish baselineThink about side-effectsAvoid trial-and-error

Know your ...

ApplicationHardwareLinux (environment)

Application

Fix bugs! aka. optimize lastRun at least -O2Multi-threadingStructure the program to be efficient on the hardware and Linuxstructure size alignment (cache-line, IO)are all POSIX/OS features used?

If in doubt test!

Hardware

Select hardware for the taskStudy the hardware you’ve selected: CPU, HT, memory, NUMA, SSD, disks, iGPU, TDP/SDPTweak the SW for the HW (think SSE, AVX)Tune HW specifics through /sys

Linux

Linux interfaces

Run up-to-date (OS+gcc+libs)/proc - more OS/sys - more HWsysctlkernel command_linekernel configtop, ntop, iotopvmstat, iostatuname

Benchmark

consistent benchmarkingcomprehensive Linux application benchmarking: Phoronix Test Suite

Profiling

perfoprofilestracevalgrind

CPU/NUMA

CoresCaches (cache line size)Layout (physical to logical mapping)AffinityScheduling

I/O

Filesystem: ext4, xfs, btrfsScheduling: noop, cfq, deadlinehdparmqueues buffersIO block sizespinning media vs SSDAIO

Memory

page sizehuge pagesmadviseswappingpage cachememory overcommitingpage faultsmemory fragmentation

Power Management

consolidate old hw -> virtualizeschedulerpowertopturn off unused hardware (isolcpus, network)scheduler (use tickless)C statesP states: frequency-voltage combinations

Networking

Throughput or latency (RT Linux?) use your HWverify link speedssegmentation offloadnetwork buffersper socket bufferszero copyTCP congestion control algorithmstools: ethtool, iperf, netstat, wireshark!

Thank You!

Questions?