1 |
On Mon, Dec 5, 2011 at 12:01 PM, James <wireless@×××××××××××.com> wrote: |
2 |
> Michael Mol <mikemol <at> gmail.com> writes: |
3 |
>> Let's start with that dual-xeon box I was using to benchmark "emerge |
4 |
>> -e @world", figure I'm looking for how better to tune my MAKEOPTS and |
5 |
>> EMERGE_DEFAULT_OPTS variables, and assume I'd like to get more |
6 |
>> information about the following factors: |
7 |
> |
8 |
> Complex and never finished, imho..... |
9 |
> |
10 |
> |
11 |
>> * What was the 1m, 5m 15m load averages? |
12 |
>> * What were the similar averages for CPU spent in user time, system |
13 |
>> time and I/O wait? |
14 |
> |
15 |
> sys-process/iotop |
16 |
> |
17 |
>> * What was network usage like? (I have a caching proxy server on the |
18 |
>> network |
19 |
> |
20 |
> Lots of different tools to look at network performance: |
21 |
> |
22 |
> wireshark, (look around /usr/portage/net-analyzer) |
23 |
> |
24 |
> |
25 |
>> so even if distfiles are lost on-system, well, a cache hit |
26 |
>> transfers at up to around 50MB/s. It'd be better, except for read |
27 |
>> performance limitations on the router box, and write performance |
28 |
>> limitations on the local machine) |
29 |
> |
30 |
> |
31 |
> bonnie++ (or bonnie) |
32 |
> |
33 |
> |
34 |
>> * What was the temperature of each CPU core, RAM module and hard |
35 |
>> drive? (Not so relevant for improving system performance, but still of |
36 |
>> interest.) |
37 |
> |
38 |
> app-admin/hddtemp (for drives) |
39 |
> |
40 |
> dunno on individual cpu cores... |
41 |
> |
42 |
>> I'd like to have a web interface I could navigate to which would show |
43 |
>> graphs of these counters. |
44 |
> |
45 |
> |
46 |
> Now all of that in one gui tool? Do post back when you get it working, |
47 |
> as I'd like to use it too!!!!! |
48 |
|
49 |
The approach I'd like to take is to have all the monitoring set up, |
50 |
launch emerge -e @world, and see what's going on around (and just |
51 |
prior to) stalls and CPU waste. I'm defining a stall as where my |
52 |
operating load falls below my number of CPU cores, and I'm defining |
53 |
CPU waste as CPU time spent anywhere but 'user'. I'd like to look at |
54 |
graphs of the metrics from over the course of the emerge. |
55 |
|
56 |
My chief thought is this: I have both 'make' and 'emerge' trying to |
57 |
reach a specific load average, which means that this particular |
58 |
dynamic system is going to have feedback as they go back and forth. I |
59 |
expect that I'll want to duck one of them under the other, but I don't |
60 |
know which one yet, and I don't know how far. |
61 |
|
62 |
I should also look to see if pbzip2 supports load awareness. Having |
63 |
eight cores suddenly start churning through BWT blocks is great if |
64 |
your load average is something like 0.24, but not so great if it |
65 |
launches your load average up to around 12. |
66 |
|
67 |
-- |
68 |
:wq |