1 |
On Sunday 30 November 2008 11:50:18 Daniel Iliev wrote: |
2 |
|
3 |
> > > real 0m26.747s |
4 |
> > > user 0m2.110s |
5 |
> > > sys 0m1.450s |
6 |
> > > localhost test # time cat test2 > /dev/null |
7 |
> > > |
8 |
> > > real 0m29.825s |
9 |
> > > user 0m1.780s |
10 |
> > > sys 0m1.690s |
11 |
> > |
12 |
> > This is not a test unfortunately. You did one run on one file and one |
13 |
> > run on another file. We do not know what else the machine was doing |
14 |
> > at that time, and that unknown is a considerable one. |
15 |
> |
16 |
> This result is from the last of three repetitions. Its values were in |
17 |
> the middle (not average). The deviations were in the range 1 to 1.5 |
18 |
> seconds. Every time reading the more fragmented was slower. |
19 |
|
20 |
That's a HUGE deviation, I assume it's real time. What's the difference in |
21 |
user and sys time? Those are much more important here. |
22 |
|
23 |
> The system was idle in level 3, no X running. I used iostat before |
24 |
> the tests and it registered several tens of KBs written for 1 min before |
25 |
> the test. Compared to the size of the test files and the speed of the |
26 |
> disk it is insignificant. |
27 |
|
28 |
You didn't remove the effect of buffers and caches. The only way I know of to |
29 |
test this correctly is to reboot the machine after every run, start no |
30 |
services and run the test with a cold cache. |
31 |
|
32 |
|
33 |
Otherwise you don't really know what you are measuring. |
34 |
|
35 |
> > Repeat your test eliminating this factor. Preferably, remount the |
36 |
> > filesystems after each run and repeat 1000 times. Then analyze the |
37 |
> > statistical distribution of your results. This should eliminate most |
38 |
> > random factors and give a more realistic real-world view. |
39 |
> |
40 |
> I'm not willing to waste time for 1000 repetitions, but why don't you |
41 |
> do the test yourself just a couple of times and see if there will be a |
42 |
> case when the more fragmented file gets read faster? |
43 |
|
44 |
You are the one making them claims! You do the test! |
45 |
|
46 |
> Actually my results are a little lower than what I expected but enough |
47 |
> for me to say that fragmentation matters. At least until proved |
48 |
> otherwise. Fragmentation leads to seeks. The average seek time on |
49 |
> modern disks is several milliseconds. Yes, there are algorithms |
50 |
> reordering the I/O requests to minimize the seek penalty, but still |
51 |
> seeks are there and they hurt performance. |
52 |
|
53 |
Buffers and cache? You have two files of 1.2G each. I have more free RAM than |
54 |
that for buffers and cache easily 90% of the time my notebook is on. |
55 |
|
56 |
-- |
57 |
alan dot mckinnon at gmail dot com |