Recomendados
Más contenido relacionado
Destacado
Destacado (20)
Similar a presentation
Similar a presentation (20)
presentation
- 1. 1Benchmarking and comparison of kernel and userspace tracers Benchmarking and comparison of kernel and userspace tracers 5 May, 2016 École Polytechnique de Montreal Mohamad Gebai Michel Dagenais
- 6. 6Benchmarking and comparison of kernel and userspace tracers Kernel tracing: the outline Final numbers: Average 90th percentile Value (ns) Overhead Value (ns) Overhead None 17 0% 17 0% LTTng 92 441% 89 424% Ftrace 118 594% 114 571% Perf 120 606% 118 594% LTTng-kprobe 122 618% 120 606% eBPF 131 671% 126 641% None LTTng Ftrace Perf LTTng-kprobe eBPF 0 20 40 60 80 100 120 140 Average latency of a tracepoint Latencyinns
- 7. 7Benchmarking and comparison of kernel and userspace tracers Kernel tracing: the outline Final numbers: None LTTng Ftrace Perf 0 50 100 150 200 250 300 Average of tracepoint latency in steady vs transient states Transient Steady Kernel tracers Latencyinns
- 11. 11Benchmarking and comparison of kernel and userspace tracers Kernel tracing: ioctl() syscall Final numbers: Average 90th percentile Value (ns) Overhead Value (ns) Overhead None 99 0% 98 0% Ftrace 311 214% 302 208% LTTng 341 244% 333 240% Perf 600 506% 586 498% None Ftrace LTTng Perf 0 100 200 300 400 500 600 700 Average latency of a syscall with tracing Kernel tracers Latencyinns
- 12. 12Benchmarking and comparison of kernel and userspace tracers Kernel tracing: open() syscall 0 10 20 30 40 50 60 0 100 200 300 400 500 600 700 800 900 1000 Latency of tracing syscall open() according to filename length LTTng Ftrace Filename length in bytes Latencyinns
- 13. 13Benchmarking and comparison of kernel and userspace tracers Kernel tracing: filtering with LTTng 0 10 20 30 40 50 60 0 50 100 150 200 250 300 350 400 450 Average latency of a tracepoint with filtering failed conditions average latency on success no filter Filter bytecode size (number of conditions) Latencyinns
- 14. 14Benchmarking and comparison of kernel and userspace tracers Userspace tracing: the outline Final numbers, but... LTTng is signalsafe reentrant LTTng is scalable Average 90th percentile Value (ns) Overhead Value (ns) Overhead None 18 0% 17 0% LW-ust 67 272% 64 276% Extrae 159 783% 118 594% Printf 245 1261% 243 1329% LTTng 255 1317% 249 1365% LTTng tracef() 536 2878% 523 2976% None LW-ust Extrae Printf LTTng LTTng tracef() 0 100 200 300 400 500 600 Average latency of a tracepoint Latencyinns
Notas del editor
- Ftrace: global clock
- Perf: difference: number of instructions (and L1 misses)
- Fast, middle, slow path of LTTng - Write - Switch page - Switch subbuffer (subbuf aligned to page size) Ftrace: - Write - Switch page (event can't be bigger than a page)
- - Extra work by the kernel - Extra work by LTTng, test the arguments, etc.
- Extrae: fast, but not flexible Event: 2 fields 1st field: event type 2nd field: entering or leaving LTTng signal-safe reentrant, atomic instructions for per-cpu variables
- Extrae: doesn't seem to be very optimised Not to take away anything from Extrae, but it seems that not much effort was put into optimizing the fast path
- - One of those tracers is not scalable - Many optimisations to be made for LTTng (get CPU id, restartable sequences for critical sections)