| 1 |
Neil Bothwick writes: |
| 2 |
|
| 3 |
> On Wed, 3 Mar 2010 12:52:55 +0100, Alex Schuster wrote: |
| 4 |
> > > The data I've seen indicates that ext2 is fastest, that's what I |
| 5 |
> > > use. |
| 6 |
> > |
| 7 |
> > I thought the small files of the portage tree especially profit from |
| 8 |
> > the notail option in reiserfs? |
| 9 |
> |
| 10 |
> They benefit compared with using reiser with tail-packing. |
| 11 |
|
| 12 |
Oh my. I have it the other way around, and never even thought much about |
| 13 |
what this does. |
| 14 |
|
| 15 |
> > Did you change the block size? |
| 16 |
> |
| 17 |
> I had to change both the block size and blocks per inode, otherwise I |
| 18 |
> would run out of inodes on a 1GB filesystem. You have to admire the |
| 19 |
> user-friendliness of ext! |
| 20 |
|
| 21 |
I only wished I could add more inodes after all are out, because this |
| 22 |
happens quite frequently to me. But yes, it's nice I can specify this at |
| 23 |
all. |
| 24 |
|
| 25 |
|
| 26 |
> > > There's no need for journalling on the portage tree, it's small |
| 27 |
> > > enough to fsck quickly and if it does get broken, reformat and |
| 28 |
> > > resync. |
| 29 |
> > |
| 30 |
> > Would the journaling overhead be noticeable? |
| 31 |
> > I also had used ext2 for my portage tree first, then I read somewhere |
| 32 |
> > that reiserfs would be the best. BTW, I have distfiles and pkgdir |
| 33 |
> > somewhere else, if not the fsck would not be so fast. |
| 34 |
> |
| 35 |
> It's certainly noticeable compared with ext3. Many benchmarks do show |
| 36 |
> ext2 to be the fastest filesystem, probably because of the lack of |
| 37 |
> journalling overhead. |
| 38 |
|
| 39 |
When I saw some, it was maybe 15% difference, and that probably due to |
| 40 |
writes I assume. The portage tree is written during sync only, and then I |
| 41 |
do not care about speed. But would accessing lots and lots of small files |
| 42 |
be slowed down by journaling? |
| 43 |
|
| 44 |
> Like you, I have $DISTDIR and $PKGDIR elsewhere, those files really |
| 45 |
> should not be mixed in with the portage tree. |
| 46 |
> |
| 47 |
> > Just for fun, I just copied my $PORTDIR into my tmpfs, emerge -DpN |
| 48 |
> > @system @world takes between 81 and 53 seconds. With reiserfs, I get |
| 49 |
> > 130 seconds first ($PORTDIR was unmounted first and mounted again to |
| 50 |
> > clear the caches), and 57 seconds in the second attempt. |
| 51 |
> > |
| 52 |
> > I had expected that tmpfs would be even faster. I think I just keep |
| 53 |
> > it the way it is now. |
| 54 |
> |
| 55 |
> The exact same thought occurred to me. With a local tree to sync from, |
| 56 |
> tmpfs seemed a good choice (you could sync it from /etc/conf.d/local) |
| 57 |
> but it seems like it is not worth bothering with. |
| 58 |
|
| 59 |
I would need more memory for that, I'm not at amd64 yet. But I probably |
| 60 |
should migrate anyway, and get another 4GB of memory. |
| 61 |
|
| 62 |
> I'll try a reiser3 |
| 63 |
> filesystem without tail packing to see if it beats ext2. |
| 64 |
|
| 65 |
I backed up my portage tree, re-created the reiserfs partition, and |
| 66 |
mounted without notail option. The same emerge command now takes about |
| 67 |
three minutes... no, on 2nd try it's five. Hmm... ah, clementine is |
| 68 |
indexing files. Why does it do this, I did not change files. Oh, and it |
| 69 |
has indexed all of my /data/mp3, while I only gave it four subfolders to |
| 70 |
index. Why does no audio player just accept my choices for what the |
| 71 |
collection is, and add other stuff? |
| 72 |
|
| 73 |
The next test gives 93 seconds, that's nice. |
| 74 |
|
| 75 |
Wonko |
Archives