1 |
Bill Kenworthy <billk <at> iinet.net.au> writes: |
2 |
|
3 |
|
4 |
> rpm is just a wrapper around a an archive with instructions on how to |
5 |
> build and or install it. I have more experience with rpm's but I |
6 |
> believe debs are the same. Just unwrap your .rpm/.deb file of choice |
7 |
> and install it manually (the binaries/code are usually in a zip inside |
8 |
> the rpm). You should in most cases also be able to get a source rpm |
9 |
> (which I suspect you are talking about anyway, but binaries do work deps |
10 |
> permitting. |
11 |
|
12 |
Yes the point is to be able to rapidly/easily be able to test and evaluate |
13 |
many cluster/hpc codes, and other codes too, before I open them up |
14 |
and look at them in detail, manually. |
15 |
|
16 |
|
17 |
> you can install rpm and then install your package via rpm - seem to |
18 |
> remember doing this with .debs somehow too. deps are a problem but |
19 |
> usually workable. |
20 |
|
21 |
This is the argument for setting up an entire RH or debian workstation |
22 |
to quickly evaluate the codes to find that less than %5 which are fast |
23 |
and worthy of a deeper look. There are dozens of 'schedulers' for clusters |
24 |
and then the concept of a 'framework' is even more loose or unconstrained |
25 |
and the myriad of associated codes (modules) that one can use. It fact many |
26 |
are pushing 'write your own framework'. It's hard to |
27 |
categorize codes into a discernible 'stack' when it comes to HPC or cluster |
28 |
offerings. Apache says one thing, a guy writing a book says another. It's |
29 |
like the wild wild west with lots of folks (mostly large corps) shoot from |
30 |
the hip and shooting off at the mouth. Many idiots and deceivers in this |
31 |
space; and I intend to benchmark, profile and expose some of the myths. |
32 |
|
33 |
So then I'll unpack the sources onto a gentoo system for deep examination |
34 |
and low level (profile ) testing and evaluation and maybe creating an |
35 |
ebuild if I like what I see. Distributing bloatware across a cluster |
36 |
or on a HPC system, is a fools errand, imho, but that is exactly what |
37 |
many are doing. It's very hard to figure out if they are 'flim_flam' artists |
38 |
or sincerely belief their own musings. |
39 |
|
40 |
|
41 |
> and why set up a workstation? - this sort of thing is tailor made for |
42 |
> vm's. Create a base for your experiments with essential packages, |
43 |
> settings etc, snapshot it (golden master) and then throwaway->restore |
44 |
> when finished with that iteration. |
45 |
|
46 |
Ultimately, I think a VM or containers environment would be keen. I |
47 |
liked docker, initially, but, like systemd, it is now bloating so much that |
48 |
it is becoming an entire operating system on it's own. Docker is too much |
49 |
for me to worry about and learn. CoreOS is still attractive to me, but |
50 |
it probably too much for me to divert into, at this time. I'm really more |
51 |
after HPC (high performance computing of scientific codes) than clustering |
52 |
of a bizzilion processes (like log/web file analytics). That sort of stuff |
53 |
is not hard to distribute, hadoop is sufficient for those routine, boring |
54 |
codes and processes that are not time_critical. I'll do some of those |
55 |
ordinary tasks but many cluster/hpc solutions already work good enough for |
56 |
routine admin style processes, imho. |
57 |
|
58 |
|
59 |
Big Science solutions using cluster or HPC are still very fragmented |
60 |
if you look at some low level performance data. It takes a very specialized |
61 |
focus to solve a given category of Big Science codes if you want to approach |
62 |
any sort of reasonable robustness. It's like you need |
63 |
a custom designed (HPCC) High Performance Computational Cluster with |
64 |
a unique (at least highly customized and tweaked) implementation for each of |
65 |
the many different possible configuration solutions. This part is evolving |
66 |
on a daily basis, so apologies if it seems sketchy because it is sketchy at |
67 |
best. I think where I'll end up is a myriad of HPCC, architecturally unique, |
68 |
running along side of each other, mostly asleep until needed. It's even more |
69 |
'fluid' when you consider the GPU resources, arm64 processors and the new |
70 |
integrated CISC chips like the amd APU and the Intel+fpga chips. |
71 |
|
72 |
|
73 |
> There are package managers besides gentoo/portage that can do a source |
74 |
> build/install and track the files on gentoo - though portage will not |
75 |
> know about it (rpm is one :) |
76 |
|
77 |
Hmmmm. Where do I read more about this? |
78 |
|
79 |
|
80 |
> and lastly, what do mean error prone? - to me a manual install is the |
81 |
> ONLY way you can build a proper ebuild that catches most of the |
82 |
> problems. In the (admittedly few) ebuilds I have done an occasional |
83 |
> human error is nothing compared to system problems for a difficult |
84 |
> package. |
85 |
|
86 |
Error prone with manual steps, due to the large amount of codes I |
87 |
am evaluating. Once I get down to less than 5% then the manual |
88 |
processes are fine, if not superior, as I'm also reviewing the codes |
89 |
of interest. It's also me, if I do not impose self_structure, I get |
90 |
sloppy with trying to manually speed things up. Also, if you go |
91 |
back to the days of .configure and look at many different make files, |
92 |
there is little coding consistency among different codes; boost and other |
93 |
lib codes are the best you can hope for on consistency, imho. ymmv. |
94 |
|
95 |
OK? Keep in mind that I'm an EE, hardware guy that started with discrete |
96 |
transistors, native assemble and state machines; and C was a luxury, so my |
97 |
perspectives do not often 'jive' with the entire OO world, but, I am trying |
98 |
to get along....... [1] |
99 |
|
100 |
> BillK |
101 |
|
102 |
James |
103 |
|
104 |
[1] http://www.phy.duke.edu/~rgb/Beowulf/c++_interview/c++_interview.html |