| 1 |
Billy Holmes wrote: |
| 2 |
|
| 3 |
> This makes it easier and faster for the kernel to context switch |
| 4 |
> between processes (and threads within a process). |
| 5 |
> |
| 6 |
> A prime example of this is a test from the LKML, where Ulrich Drepper |
| 7 |
> wrote he was able to start and stop 100,000 threads in about 2 |
| 8 |
> seconds. In further messages, Ingo Molnar confirms this and also |
| 9 |
> points out that before these changes, it would have taken upwards to |
| 10 |
> 15 minutes for the same thing. |
| 11 |
> |
| 12 |
> http://kerneltrap.org/node/422 |
| 13 |
|
| 14 |
Absolutely true. NPTL is drastically more scalable than linuxthreads. |
| 15 |
For instance, spawn 1000 worker threads in a JVM and the scheduling |
| 16 |
overhead alone will cause the system to become unresponsive across the |
| 17 |
board. Use NPTL and you can get 10000 threads all running no problem. |
| 18 |
|
| 19 |
-ryan |
| 20 |
-- |
| 21 |
gentoo-server@g.o mailing list |
Archives