F@H - 18K PPD

[KE1HA]

Just upgraded all my boxes to the latest Linux SMP client, added 3 more clients to each Quad and 1 more to my E6800. I find it hard to belive that it will bring my PPD to roughly 18K. I was bringing in about 10.5 to 11K PPD.

Also, it looks like I need to pump up the volume on my first Quad box, as it's not performing as well as the second.

Does this Seem Right?

.

 

[rich99million]

If you set FahMon to estimate the PPD based on the last 3 frames or last frame only that should give you a decent idea of the actual PPD - I think by default it's set to average all frames in which case your single-client times will be majorly skewing the numbers.

I'm not totally sure about the need for 4 clients personally, though Cob swears by it and he does alright for himself (though I think he uses VMWare which needs 2 instances to be able to access all 4 cores in the first place and two clients for each VM to max out the CPU). When running natively I'd hope it wasn't really necessary to use more than 2 clients to max out the CPU (the Linux client used to leave ~10% idle) though if people's results show otherwise I guess it could be worth it.

 

[Mattus]

If you set FahMon to estimate the PPD based on the last 3 frames or last frame only that should give you a decent idea of the actual PPD - I think by default it's set to average all frames in which case your single-client times will be majorly skewing the numbers.

Yeah, those numbers look a little out... your Q6600s are producing 5000-6000ppd each, which isn't possible even with those specs!

Nice output though

 

[KE1HA]

If you set FahMon to estimate the PPD based on the last 3 frames or last frame only that should give you a decent idea of the actual PPD - I think by default it's set to average all frames in which case your single-client times will be majorly skewing the numbers.

Thanks Rich, I wasn't aware you could change the sample range in FAHMon. Will have a look at that and see how it changes things.

Im going to try this for a week and see what the real WU & points count are then make the call whether to run 1 or 4 clients.
.

 

[SiriusB]

I have been running 4 clients natively for a fortnight now. I started with around 2K PPD, I am pushing 4K now.

You should probably trust FahMon more than EOC, as if you go a long time without dumping your PPD goes right down. Less likely for you, but after a mass EUE all my clients are all at the same % - so I only dump points once a day or so. [albeit almost 6K

 

[KE1HA]

There's definately a difference in PPD when changing the calculation frequency on FaHMon:

Ignore Asynchronous Clocks = Not Selected
Prefrences {All Frames} = 17,975 ppd
Prefrences {Last Frame Only} = 17,250 ppd
Prefrences {Last 3 Frames} = 12,950 ppd

Ignore Asynchronous Clocks = Selected
Prefrences {All Frames} = 9,628 ppd
Prefrences {Last Frame Only} = 17,774 ppd
Prefrences {Last 3 Frames} = 12,952 ppd

Basically, it's inconclusive. Both methods show similar results in {Last Frame Only} and {Last three Frames}, with Synchronous and Asynchronous skewing the data in extreme's both ways.

I will know soon enough, as all the first WU's are nearing their completions Times

Next 6 Hours: 6400
Next 12 Hours: 5807
Next 18 Hours: 4430

Tota in 18 Hours = 16,637

Not to Bad for 1.125 days, but it does suggest that the 24 hour ppd is closer to 12K to 13K v.s. the 18K, which begs the question why would one want to run 4x clients v.s. 1x.

I suppose, that if your a Beta WU Tester or have concerns over EUE's due to OC (would not apply to network issues obviously) then the risk is lower going with 4x clients, as you would only loose produciton on one WU (hopefully) in an EUE or Beta WU malfunction situaiton, as opposed to loosing your entire production from a particular node when running only one client.
.

 

[Cob]

I would let them run for three or four days before drawing any conclusions.

I did a fair bit of testing in native Ubuntu and found that four clients was the best course.

The best a single client could manage was something like 7-8mins per frame, whereas four running together could manage 19-20mins each, giving an average of ~5mins per frame.

 

[KE1HA]

That's the Plan Cob ... going to let them run a week or so and see what's really going on.

Here's the latest. I reinstalled everything yesterday. ASUS-1 & ASUS-2 are still running One Core Each.

.