1 |
Apparently, though unproven, at 02:56 on Thursday 09 September 2010, Enrico |
2 |
Weigelt did opine thusly: |
3 |
|
4 |
> * Alan McKinnon <alan.mckinnon@×××××.com> wrote: |
5 |
> > > True. But FreeBSD isn't that popular like Windows, Mac or Linux. |
6 |
> > |
7 |
> > So you don't work at a Tier 1 ISP then? |
8 |
> > |
9 |
> > FreeBSD rules that space. I get hugely better performance out of Postfix |
10 |
> > on FreeBSD than on Linux - all other ISPs in this country concur. |
11 |
> |
12 |
> Well, not everybody is a tier-1 isp ... ;-o |
13 |
> |
14 |
> BTW: one of my customers, a really big one here in Germany |
15 |
> (who also has several of the major free mail portals) runs |
16 |
> its mail systems on GNU/Linux (well, inhouse mailing is done |
17 |
> via Exchange+ADS, surprisingly it actually works ;-)). |
18 |
> |
19 |
> But I'd really like to know what produces the performance hits |
20 |
> on Posfix @ Linux. |
21 |
|
22 |
It comes down to the IO scheduler. Linux is designed to be general purpose. |
23 |
FreeBSD is designed to be much more specific. |
24 |
|
25 |
Both are very good at what they do, the trick is in realising what those |
26 |
things are and playing to their strengths. |
27 |
|
28 |
|
29 |
> > In fact, portage is complete overkill and I refuse to allow it |
30 |
> > to be deployed at work. Check my posting history for the |
31 |
> > rationale behind this. |
32 |
> |
33 |
> Well, portage could be much thinner if certain things would be |
34 |
> moved explicitly out-of-scope or solved more generic on a |
35 |
> different layer. (yes, I'm explicitly ignoring the historical |
36 |
> issues right now ;-p). |
37 |
|
38 |
My beef with portage in my specific production setup is the amount of work it |
39 |
takes my guys to keep everything up to date. We don't have 150 identical |
40 |
servers in a farm (I'd love that and would switch to Gentoo immediately if it |
41 |
were). I have 130 completely different configs and uses for those servers. |
42 |
|
43 |
The maintenance admins would have to fully grok all of portage, the |
44 |
implications, predict the outcome and understand what they are about to update |
45 |
every time they do an update. And they'd have to know it for 130 permutations. |
46 |
|
47 |
Heck, *I* can't track that, I won't expect someone else to. Centos better |
48 |
suits our needs - deploy X, you know what you are going to get. Having said |
49 |
that, all my personal stuff and my own dev boxes are Gentoo. Why? Coz I can |
50 |
change stuff around for testing on a whim, figure out the right approach then |
51 |
document what to do on Centos to achieve that. |
52 |
|
53 |
|
54 |
> |
55 |
> For example: |
56 |
> |
57 |
> * distro-specific and various source retrieval methods would not |
58 |
> be necessary, if the packaging/distro-build system would simply |
59 |
> fetch it's sources from an vcs (eg. git ;-p) using canonical |
60 |
> versioning/namespace scheme [1]. |
61 |
> |
62 |
> * instead of useflags (the terminology implies we're switching |
63 |
> things some package *uses*, not provides), model the available |
64 |
> features, eg. like Briegel [2] does. (that's more a methological |
65 |
> that a technical issue). |
66 |
> |
67 |
> * instead of slotting, assign separate package names when multiple |
68 |
> version concurrency is required (and maybe pull them together |
69 |
> via virtuals) |
70 |
> |
71 |
> * rely on an pure DAG as dependency graph - per definition. |
72 |
> when circular dependencies occour, fix them in the source tree, |
73 |
> for example splitting off certain packages in several smaller ones. |
74 |
> |
75 |
> |
76 |
> [1] http://www.metux.de/download/oss-qm/normalized_repository.pdf |
77 |
> [2] https://sourceforge.net/p/briegel/ |
78 |
> |
79 |
> |
80 |
> Don't get me wrong, that shall not be understood as ranting against |
81 |
> Gentoo, just showing suitable approaches we'd start afresh on a |
82 |
> "green grassland" (w/o all the historical burdens). |
83 |
> |
84 |
> |
85 |
> cu |
86 |
|
87 |
-- |
88 |
alan dot mckinnon at gmail dot com |