1 |
On Fri, 5 Dec 2003 09:43:13 -0800 Bob Miller <kbob@××××××××××.com> |
2 |
wrote: |
3 |
| > On Thu, 04 Dec 2003 19:26:51 -0600 Steven Elling <ellings@×××××.com> |
4 |
| > wrote: |
5 |
| > | P.S. Developers: You might want to look into adding a number to |
6 |
| > | the'-j' option in MAKEOPTS during emerges of certain packages to |
7 |
| > | account for this situation. |
8 |
| > |
9 |
| > Wouldn't that be rather unfair upon users who have a 100 CPU distcc |
10 |
| > cluster? |
11 |
| |
12 |
| My testing showed that distcc doesn't seem to scale beyond about 8-10 |
13 |
| hosts. (Hosts, not CPUs.) Since the originator has to do all the cpp |
14 |
| processing and all the linking, the other hosts start standing around |
15 |
| idle when the originator reaches 100% CPU. |
16 |
|
17 |
It depends a lot upon what the central host is. If the central host is |
18 |
faster and has several CPUs (for example, a server using workgroup |
19 |
machines as slaves) then distcc can potentially scale very well. If the |
20 |
central host is a really old sparcstation5 then adding more than about |
21 |
two slaves doesn't help. |
22 |
|
23 |
It also depends upon the makefile in question. Very few Makefiles allow |
24 |
more than maybe 20 things to be compiled simultaneously anyway. |
25 |
|
26 |
So maybe I should rephrase... |
27 |
|
28 |
Wouldn't that be rather unfair upon users who have a dozen 8-way amd64 |
29 |
boxes fully interconnected by switched gigE who do glibc development? |
30 |
|
31 |
-- |
32 |
Ciaran McCreesh |
33 |
Mail: ciaranm at gentoo.org |
34 |
Web: http://dev.gentoo.org/~ciaranm |