1) Message boards : Number crunching : Boinc farms. (Message 15147)
Posted 20 Oct 2006 by YeshuaAgapao
My Alienware laptop has a desktop Pentium 4 in it. It is 2.9A AC-In with a 300W requirement for inverter for a car dapter. That is 180 W (20V, 9A) DC-out and 300W AC-In (probably actually 100V @ 2.9A = 290W). 180/290=0.621 - 62.1%. So Consumer devices (excluding ultra-light travel adapters) are probably really 50-70% efficient. Using the same pattern for Dell - 1.5A @ 100V (150W) in, 20V 4.5A out (90W) - 60% efficiency. My Dell Inspiron has a Penitum 4-M.


That's not correct.

Typical (modern) Power Supplies operate at approx. 70-85% efficiency (where >80% is very good). In the past, these efficiently levels were known to be only reached unter ~75-80% load conditions, something which has recently changed to the better.

A 300W PSU can never pull more than ~300W from the Plug (exact specs and safety circuits usually cut in slightly below that, depending on which supply lines hit their individual Ampere limit first).
It is a common error to believe e.g. a 500W PSU would actually draw 500W of power all the time. In fact, a good 500W PSU can operate at a sustained 40W with a power-saving Hardware setup attached to it. The attached Hardware is what defines the power consumption (minus a factor of the PSU's efficiency).

I'm running modern Systems with way overpowered PSU's (300-420W, since it's sheer impossible to get e.g. a 150W ATX 2.0 PSU nowadays), and they're effectively using 75W to 115W in total (depending on Hardware Setup) from the Plug (measured).

A good Laptop with crunching in mind should not use more than 45-50W unter full CPU load.
Something in the figures of 100W or more are power-wasting Desktop replacement Notebooks and terribly inefficient when crunching comes into play; 180W for any Laptop would be completely unacceptable. (A well setup crunching Desktop uses far less than that and offers higher performance at the same time)

A Pentium4 is about the worst-case scenario (highly inefficient and by now comparably slow when compared to other architectures)... The only CPU to be less efficient than the P4 was the P4 based Celeron (which marks the rock bottom of poor efficiency in the whole x86 field).

2) Message boards : Number crunching : Boinc farms. (Message 15072)
Posted 12 Oct 2006 by YeshuaAgapao
Transformers for consumer devices are 40-60% efficient. So you basically double the power supply's capacity to find out how much juice will be sucking in. A 300W power supply will suck in 600 watts. A 400W PS on a gaming rig will suck in 800 watts from the AC outlet. BOINC will cause the PC to consume about 70-80% of the power supply's rating, 90-100% if you have a lot of drives or are pushing the video card with your screensaver or game.

600 watts at 10c/KW-h (SRP is 10.8c - 8.8c base + 2c 'fuel price surcharge') is 6c/hour, $1.44 per day, $43.20 per month!!!. So if you are running BOINC on a too old of a machine you are wasting money to your electric company. Is putting an old 100-400mhz desktop worth $40/month in electricity costs (assuming you run it with BOINC 24/7)?

Laptops are far more power efficient, taking in 140 (standard laptop) to 180 (gaming capable with mobile components) to 300 watts (gaming rig with SLI and desktop processor), but you will need an industrial-office fan to keep them cool. Alienwares can fend for themselves because they got 4 fans, just left it off the desk a little. Dells and Toshibas need help. The big 30" patton fan costs $200 to buy and uses 3amps of power, 1/2 of a desktop, or 1 laptop,and it can service 4-8 (depnding on how big/flat your desk is!) machines and a human or two. So if you want to run a geeky BOINC farm, I suggest using old laptops instead of old desktops. You will notice the difference in the electric bill, about $20 per machine.
3) Message boards : Number crunching : I think we should restrict work units (Message 14433)
Posted 25 Jul 2006 by YeshuaAgapao
I think they are 'restricting' the workunits from the hog by doing smaller batch sizes. They run out of work in just an hour or two. Big Cache helps more now. It knocked two of my computers out because one has a internet connection that has to compete with Emule P2P (because I'm not getting rid of it) and the other that is only online when I swap the vzw pc-card to it. So only my work computer will ever get work now because it has a internet conncection that is good for it begging for work every few minutes.
4) Message boards : Number crunching : 2 WUs won't upload (locked by file_upload_handler PID=20350) (Message 14061)
Posted 20 Jun 2006 by YeshuaAgapao
Here's the error. Its been retrying for about 10min. Is this a temporary thing or are they bad workunits? I had 7 other WUs upload ok.

6/19/2006 8:06:07 PM|LHC@home|Started upload of wmay1C_v6s4hvnom_mqx-hlf__7__64.303_59.313__6_8__6__15_1_sixvf_boinc270513_0_0
6/19/2006 8:06:07 PM|LHC@home|Started upload of wmay1C_v6s4hvnom_mqx-hlf__19__64.294_59.304__8_10__6__10_1_sixvf_boinc286237_3_0
6/19/2006 8:06:12 PM|LHC@home|Error on file upload: [wmay1C_v6s4hvnom_mqx-hlf__7__64.303_59.313__6_8__6__15_1_sixvf_boinc270513_0_0] locked by file_upload_handler PID=20313
6/19/2006 8:06:12 PM|LHC@home|Error on file upload: [wmay1C_v6s4hvnom_mqx-hlf__19__64.294_59.304__8_10__6__10_1_sixvf_boinc286237_3_0] locked by file_upload_handler PID=20350
6/19/2006 8:06:12 PM|LHC@home|Temporarily failed upload of wmay1C_v6s4hvnom_mqx-hlf__7__64.303_59.313__6_8__6__15_1_sixvf_boinc270513_0_0: transient upload error
6/19/2006 8:06:12 PM|LHC@home|Backing off 1 minutes and 12 seconds on upload of file wmay1C_v6s4hvnom_mqx-hlf__7__64.303_59.313__6_8__6__15_1_sixvf_boinc270513_0_0
6/19/2006 8:06:12 PM|LHC@home|Temporarily failed upload of wmay1C_v6s4hvnom_mqx-hlf__19__64.294_59.304__8_10__6__10_1_sixvf_boinc286237_3_0: transient upload error
6/19/2006 8:06:12 PM|LHC@home|Backing off 1 minutes and 10 seconds on upload of file wmay1C_v6s4hvnom_mqx-hlf__19__64.294_59.304__8_10__6__10_1_sixvf_boinc286237_3_0

5) Message boards : Number crunching : I think we should restrict work units (Message 13982)
Posted 14 Jun 2006 by YeshuaAgapao
I not only had 2 days of work left when LHC ran out of work, I also suspend all other projects on all machines in the duration that LHC has work to offer, then turned them back on when it ran dry, so it takes mee 4-5 days to clear the cached WUs out.

The INCA computer (the 3.4ghz with 2GB RAM) is not always connected so that one has the worst lagtime on WU-return. That machine i grub it all up from all projects (I only connect it regularly when LHC has work) and wait until 1 day before the shortest deadline to connect it again. Thats why that machine has the most CPDN and CPDN-Seasonal credits. INCA is 10-days on all projects (home profile). The other 2 machines are work profile (1-day cache) on all projects but home for LHC for expert scarce-WU pigging. The 1.7GHZ machine is my file sharing (Emule) machine and it keeps the Verizon Wireless CellPhone Modem (I'm 99.999% bandwidth hog on VZW using 30-60GB per month on a traffic-shaped 14Kup/14down low prio - 5-7K daytime - bandwidth) . The 1GB 3.4MHZ is me using BOINC at work.

If LHC always had work then I wouldn't be so greedy :) And I'd probably be at 85th percentile for RAC/Total and not 99.5% RAC and 96% total :) Check out this: I pigged about 4600 credits in the past 3 work batches

It is important to MaxCache LHC becuase it takes only 18-30 hours for them to run out of work once they put it up (looks like they put up 80000-150000 WUs at a time).

What's important is getting the work done. As I write this, GreatInca, one of your computers still has eight results left to run, while my computers have been out of work for two or three days. Had you not filled your cache to excess the eight results could have been assigned to other computers and the work completed days ago. Spread this across thousands of users and the tail of unreturned results drags out for days.

Each study generally requires analysis of the previous study to determine appropriate parameters. If each study is completed more quickly the next one can be released sooner. The message, therefore is:

Keep your cache as small as possible.

6) Message boards : Number crunching : Work to be done! (Message 13822)
Posted 2 Jun 2006 by YeshuaAgapao
The cache size only is usefull once LHC is freshly out of work. You can set cache size for a specific project by create separate preferences for one of your home/school/work categories. As computers are assigned to these categories on a per-project basis, you can say set CacheSize to 1 for default, and to 10 for 'school', and then for LHC@Home, you can put all you computers as 'school', and 'work' or home' for all the other projects, so you can get a 10-day cache only for LHC@home, and 1 day for your other projects. LHC's short deadlines greatly decrease the usefulness of the large caches though. It won't get 10 days worth of work. 3-4 days at most (if none bomb-out early from beam destabilization).

Thats where the next trick comes in - suspend all pronects except LHC and crunch LHC exclusively while and only while LHC has work to serve. You will fetch and cache LHC work as if LHC was your only project. Plus you get maximum throughput while there is work to fetch. Once LHC runs out of work, turn back on all the other projects and BOINC will finish all the LHC units at its own leisure. Heavy multi-project people will be in NDF mode though, but you still get the large post-work exhaustion cache.
7) Message boards : Number crunching : I think we should restrict work units (Message 13724)
Posted 24 May 2006 by YeshuaAgapao
It is important to MaxCache LHC becuase it takes only 18-30 hours for them to run out of work once they put it up (looks like they put up 80000-150000 WUs at a time). I seem to be getting an edge by suspsending all projects but LHC not just to get work (until the downloads start), but for the entire 18-30 'work window'. Then you turn all the other projects back on and let BOINC take its time. So you crunch LHC exclusively while theres more work to serve, but after that it doesn't matter.

So you crunch just LHC for the 18-30 hours and then when it goes 'out of work', you still have the maximum possible cache as if LHC was the only project registered in BOINC and you turn all the other projects back on (LHC-exclusive is pointless when theres no more work left to serve).
8) Message boards : Number crunching : Is the user base/project participants growing a bit too large, for our server? (Message 13433)
Posted 24 Apr 2006 by YeshuaAgapao
LHC hardly ever has work. If LHC has work you'd better grub it up. I used to not like LHC's short deadlines and BOINC's behavior with it but I didn't know the LHC's work was very intermittent. Now i like it. To grub up the most possible work, make a home/work/schoool profile (probably school) that has 10 day cache and assign all your compmputers to that special profile for LHC only (Global prefs but profile assignment is per-project). Then Suspend all projects but LHC and press update and you should grub up to your WU quota in LHC. Once downloading is finished, turn all your other projects back on. About 3-4 days of non-stop crunching. Your CPDN WUs will guarentee BOINC going to NDF mode (immediately upon download) and crunching LHC first (and well before the deadline). You may get 2 trips before LHC's work batch runs out if the machine is fast enough.
9) Message boards : Number crunching : I think we should restrict work units (Message 13380)
Posted 15 Apr 2006 by YeshuaAgapao
My cache is set to MAX on all 3 machines for LHC because it usually don't have work and it takes less than 2 days for it to run out and 1 machine is usually not connected (alienware laptop), my old dell is on crappy internet (semibraodband cellphone modem competing with file sharing), and the other one is at work so its got a good connection. LHC is at 15% for me but its struggling to reach 10% of my total credit because it don't always have work. Oh yeah one trick to force your boinc to dowload the most possible LHC workunits is to suspend all other projects, update on LHC, then resume all the other projects. Sometimes on a one or two day window when LHC has work, boinc dont want to download any. THe suspend-download-resume trick works really well and LHC's short deadlines means they get crunched first, and I don't care if I can't complete all the QMC work on time.

On all other projects except the CPDN ones only the alienware laptop is set to MAX. I also shut down (suspended) QMC because they lie to BOINC about crunch times, making your boinc crunch nearly 2 weeks straight on NDF mode even on a non-maxCache machine (their workunits take 40 hours not 15 or 20). I might turn QMC back on in a few weeks but if they send 8 40-hour work units again i probably will abort most of them and only crunch 2 or 3 of them.

Ufuids is having problems and is only serving a masters thesis. Their database crash made me lose out on credit for 12 WUs - 5+2+5 for each machine - and they take 15 hours each to crunch.

©2024 CERN