This video was mentioned on highscalability.com, so I thought I'd have a look. Knowning this stuff is useful when you're in the business of delivering large amounts of geographical data to a large amount of clients.
C10M = 10 million concurrent requests.
In short the video is about getting a server to handle 10 million concurrent requests, and how the problem isn't hardware. It's software (e.g. Apache, UNIX etc).
Discussions related to the video:
- c10m.robertgraham.com (presenters blog in the topic)
- The Secret To 10 Million Concurrent Connections -The Kernel Is The Problem, Not The Solution (High Scalability)
- Is It Time To Get Rid Of The Linux OS Model In The Cloud? (High Scalability)
- Hacker News thread
A C program shown in the talk:
#include <stdio.h> int main() { char *ip1 = "10.1.2.3"; unsigned char ip2[] = {0xA, 0x1, 0x2, 0x3}; int ip3 = 0x0A010203; int ip4 = *(int*)ip2; printf("ip3 = %x\n", ip3); printf("ip4 = %x\n", ip4); return 0; } |
Some take aways (Roberts take-aways at 44m40s):
- The kernel is not the solution. It is the problem
- Google "lock free" data structures
- You can configure Linux so use only a fixed subset of the processors in multi-core:
maxcpus=2(boot param to make Linux only use the first two cores), i.e. you own the rest. - You can configure which core your thread should run on:
pthread_setaffinity_np() - You can configure which cores handles which interrupts:
/proc/irq/smp_affinity - Following a pointer => cache miss. In a high scalability network server there is budget for four cache misses (300 clock cycles per miss). This means, colocate data in a contiguous memory data structure