1 |
On Sunday 27 May 2007, Christopher Friedt wrote: |
2 |
> Using the stage3 gentoo arm-softfloat-linux-uclibc filesystem, I had for |
3 |
> some time been natively compiling packages for my target system, with |
4 |
> the root located over NFS. |
5 |
> |
6 |
> Normally it wasn't so bad - sure it was quite slower than |
7 |
> cross-compiling. But one thing I hated about cross compiling was |
8 |
> randomly running into packages that use some sort of pkgconfig file that |
9 |
> I didn't have present, or trying to run some native program in configure |
10 |
> and coming up with an error. |
11 |
|
12 |
ive found NFSv3 + TCP actually isnt so shabby ... you have to make sure to use |
13 |
those options in the kernel cmdline though as the default is NFSv2 + UDP |
14 |
|
15 |
> When I got around to compiling the GNU Classpath (for use with jamvm, |
16 |
> and stripped of all but the necessities), I found that compiling this |
17 |
> app over an NFS root failed quite often because of out-of-memory errors. |
18 |
> The process was killed outright. Then I read somewhere about creating a |
19 |
> swap file on an NFS root, using losetup to create a loop device, and |
20 |
> using swapon to load that swap into the kernel's VM. That turned out to |
21 |
> be the worst idea I think I've ever had, because the overhead of using |
22 |
> RPC with the NFS root was probably of O^2 complexity at that point, |
23 |
> especially considering all of the swapping that had to be done with such |
24 |
> a huge compile, using jikes of course. |
25 |
|
26 |
this is the biggest problem and ive found not worth the hassle if possible ... |
27 |
doing network swap (whether it be via nbd or nfs or ...) is just not usable |
28 |
|
29 |
if your machine doesnt have enough physical ram to compile w/out -pipe, then |
30 |
it's time to search for another option :) |
31 |
-mike |