HostID 1001562 - Richard Haselgrove's Q6600 Quad Core

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109409777859
RAC: 35113633

RE: Yes please, if it's

Message 78887 in response to message 78885

Quote:
Yes please, if it's easy it would be helpful to have them sorted ...

OK, I've made the change and to test it I went looking for another host on which I could do an "initial grab" :). I think I've found an interesting candidate to add to the group.

I've really enjoyed pondering the plots you have presented so far and reading Bikeman's explanation of the wiggles. The troughs seem to be well categorised now and I'm just waiting to see some peaks. So I've chosen a host whose very latest data sweeps right through a peak and a bit down down the other side. There is a good range of other seq#s and slightly different freqs as well. There are a couple of app transitions - the latest being 4.27 to 4.35. It's also an AMD host to add to the interest.

I've just uploaded the initial grab and you can find it here. It seems to be sorted properly.

Cheers,
Gary.

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109409777859
RAC: 35113633

RE: I've just uploaded the

Message 78888 in response to message 78887

Quote:
I've just uploaded the initial grab and you can find it here. It seems to be sorted properly.

There's something wierd with the very latest data for this host. Two days ago, crunch times were around 25K seconds - admittedly down towards the trough. The RR_V7E plot for 14 data points in the 822.25 freq band look OK. From those points the peak is estimated at around 38K secs.

More recently at the 822.60 freq band, there are 31 points spanning a peak. RR seems to struggle a bit with these and it's little wonder. The peak runtime is around 65K secs with very sharp drops on either side of the peak. The model can't get up to the peak runtime and in the process has dropped the predicted trough to way below where previous information indicated it should be. The variance figure came out at 0.876, a factor of three or four higher than you would otherwise expect.

My initial guess is that there is something happening on the host which is making the data useless, unfortunately.

Cheers,
Gary.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2139
Credit: 2752987842
RAC: 1376008

RE: I've really enjoyed

Message 78889 in response to message 78887

Quote:

I've really enjoyed pondering the plots you have presented so far and reading Bikeman's explanation of the wiggles. The troughs seem to be well categorised now and I'm just waiting to see some peaks. So I've chosen a host whose very latest data sweeps right through a peak and a bit down down the other side. There is a good range of other seq#s and slightly different freqs as well. There are a couple of app transitions - the latest being 4.27 to 4.35. It's also an AMD host to add to the interest.

I've just uploaded the initial grab and you can find it here. It seems to be sorted properly.


Interesting find. There weren't enough data points for the very earliest app, but plenty for 4.27 and 4.35, so I've done them as two separate plots. I've thrown away a few sequence numbers above 500, but we've got good fractions of a cycle for both apps. He's even given us a bit of an overlap at 821.85.

App 4.27

(direct link)

App 4.35

(direct link)

In spite of your concerns about the data, that looks like a very shapely plot to me - but as you say, look at that left-hand time axis. Very odd. I think that might be something to report in the 'Power Users' thread - but he is using some fairly unusual equipment: I make it 8 dual-core Opterons (16 CPUs total), and approaching 48GB of RAM?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109409777859
RAC: 35113633

I've uploaded the latest data

Message 78890 in response to message 78882

I've uploaded the latest data for the 6 hosts shown below.

1001562 belonging to RH.
997488 belonging to peanut.
1100395 belonging to smoked trout.
1101980 belonging to RAMA.
1106990 belonging to Herr Datenrat.
989926 belonging to Elphidieus.

For some reason, nivekfalk's host still isn't returning any new results so I've temporarily dropped it until it starts working again.

I've added the Elphidieus host since it's returning lots of results and I'd like to see what else will turn up that's unusual or unexpected :). It has some nice multicycle data for the 4.10 app which looks very good. There is also enough 4.33 data to give peaks and troughs but the RR plot is rather wierd. You'll have to look for yourself as it defies reasonable description :). What is immediately unusual is that there are 166 results all for the 865.85 freq whose seq#s lie between 282 and 447 - ie exactly 166 seq#s. I'd better go look more deeply as my impression is that the scheduler doesn't allow you to get that many without a freq change. Also it's the magical xxx.85 that I seem to recall someone reporting something unusual about previously :).

EDIT:
I didn't notice it before but the nice plot for the 4.10 app has a freq of 865.85 as well - 66 data points. The mixed results are also 865.85. There seem to be a grand total of 338 data points all from the same frequency band. There are many other freq bands in the data with very few points - maybe 10 or more bands with about a total of 30 points or so. It seems very unusual.

Cheers,
Gary.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2139
Credit: 2752987842
RAC: 1376008

RE: I've uploaded the

Message 78891 in response to message 78890

Quote:

I've uploaded the latest data for the 6 hosts shown below.

1001562 belonging to RH.
997488 belonging to peanut.
1100395 belonging to smoked trout.
1101980 belonging to RAMA.
1106990 belonging to Herr Datenrat.
989926 belonging to Elphidieus.


I'll see what I can plot up later today - got to do some real work first :-(

Quote:
I've added the Elphidieus host since it's returning lots of results and I'd like to see what else will turn up that's unusual or unexpected :). It has some nice multicycle data for the 4.10 app which looks very good. There is also enough 4.33 data to give peaks and troughs but the RR plot is rather wierd. You'll have to look for yourself as it defies reasonable description :). What is immediately unusual is that there are 166 results all for the 865.85 freq whose seq#s lie between 282 and 447 - ie exactly 166 seq#s. I'd better go look more deeply as my impression is that the scheduler doesn't allow you to get that many without a freq change. Also it's the magical xxx.85 that I seem to recall someone reporting something unusual about previously :).


I was having a play around with th3's host 1109009, because he mentioned in another thread how fast 4.35 is - I thought it might throw some light on Herr Datenrat's extreme cyclic variation. Well, th3 seems to be showing no cyclic variation at all - just random noise, but with one run almost 50% longer than the rest. Gary, Mike - you might like to take a look.

th3
th3
Joined: 24 Aug 06
Posts: 208
Credit: 2208434
RAC: 0

RE: I was having a play

Message 78892 in response to message 78891

Quote:
I was having a play around with th3's host 1109009, because he mentioned in another thread how fast 4.35 is - I thought it might throw some light on Herr Datenrat's extreme cyclic variation. Well, th3 seems to be showing no cyclic variation at all - just random noise, but with one run almost 50% longer than the rest. Gary, Mike - you might like to take a look.

I was planning to make a post about that long WU when the wingman finishes, could take a few days, he still got 5 WUs to go before he will start on that one.
I noticed that WU already while it were running so i had a check to rule out the possibility of thermal throttling, found nothing, temp was as usual and the WU running simultaneously was within the normal variation.

This one

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109409777859
RAC: 35113633

I've uploaded the latest data

Message 78893 in response to message 78890

I've uploaded the latest data for the 9 hosts shown below. I'm not intending to keep expanding this exponentially, BTW :). The nivekfalk host is back with a vengeance, there's RH's new host and also the host of th3 that RH mentioned. This is all in addition to the ~100 hosts of my own that I'm also collecting data for :).

1001562 belonging to RH (#1).
1001564 belonging to RH (#2).
997488 belonging to peanut.
1100395 belonging to smoked trout.
1084360 belonging to nivekfalk.
1101980 belonging to RAMA.
1106990 belonging to Herr Datenrat.
989926 belonging to Elphidieus.
1109009 belonging to th3.

I haven't had time to look at any of the above in any depth. With the rash of new apps for both Windows and Linux, I'm having difficulty just keeping my fleet even partially up to date.

PS:
There is no new data for the smoked trout host. I've just checked and there is no further work in the cache so I guess that might be one to remove.

Cheers,
Gary.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2139
Credit: 2752987842
RAC: 1376008

RE: I'm not intending to

Message 78894 in response to message 78893

Quote:
I'm not intending to keep expanding this exponentially, BTW :).


Thank goodness for that! I'll try and get something plotted: visually, 4.36 is even faster than than 4.32.

Lost a bit of time to a power outage yesterday (gales), but no permanent damage: and the computers crunched unperturbed through the earthquake, though I felt it myself. Just waiting for the plague of frogs ;-(

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2139
Credit: 2752987842
RAC: 1376008

For today's viewing pleasure,

For today's viewing pleasure, may I present:

Peanut:

(direct link)
Old faithful, beginning to get a clear picture of the 4.33 cycles. Just a couple of odd wiggles in 910.10

Nivekfalk:

(direct link)
Had a slow block in the middle of 864.35, but is now back on track. May have had hardware problems, or may have taken time out to run another project.

Elphidieus:

(direct link)
Plotted to show the difference between 4.10 and 4.33. Let me know which bit of the weird wiggle you want to zoom in on!

Other monitored hosts:
Smoked Trout - didn't really add anything new before he took his break.
Rama - beginning to get the general shape of the cycle, but the individual points are all over the place.
Herr Datenrat - new data generally following the curve we saw before, but nothing to confirm that extraordinary peak.
th3 - steady progress, still just that one wild point over 14Ksec.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6534
Credit: 284730859
RAC: 105773

RE: Lost a bit of time to a

Message 78896 in response to message 78894

Quote:
Lost a bit of time to a power outage yesterday (gales), but no permanent damage: and the computers crunched unperturbed through the earthquake, though I felt it myself. Just waiting for the plague of frogs ;-(


Oh, they come just after the Four Horsemen ..... :-)

Terrific plots! I was looking right in the troughs, hoping for an inspiration like - find the lowest point, then towards the right the pattern is up, down, then two ups followed by a down - but alas nothing so consistent. However this rather begs for a per-point weighting applied in the least squares method - the general idea being weighting goes inversely to 'spread' - in order to retain the overall success of the sine curve fit but yield some prediction without outright rejections. Perhaps quixotic! The worst of the wobbling is not at the minimum but in the principal value range of ~ 0.8 to 0.95 ( or beyond ~ 6 tasks either side ). Definitely a strong argument for including a 'fitness' metric in RR.

Cheers, Mike.

[ edit ] If you have a point or two well off the trough to pin the algorithm there's no issue, it's only if your data is purely 'peri-trough' that the flap problem arises.

I have made this letter longer than usual because I lack the time to make it shorter ...

... and my other CPU is a Ryzen 5950X :-) Blaise Pascal

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.