| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Community
Forum: Teams
Thread: UK |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 14623
|
|
| Author |
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
I'm confused by this thread concerning UGM and HyperThreading.
It's true that, if there's a problem running multiple WU's in parallel, switching HT off will reduce the contention. This is simply because you've switched off half the "processors". It's also true that if you switch HT off then the WUs will finish quicker. But I would question whether the level of contention is high enough to justify switching HT off, because HT is there for a good reason. HyperThreading works by using a single execution unit but two parallel instruction decode pipes. This means that if one of the pipes stalls, the other can still feed the execution unit. The overall effect is a small but useful performance increase. I have never switched HT off because I have always found that the overall throughput of the CPU (chip) is higher with it on. It might only be a few percent, but it's always been true. Any contention which causes the instruction queue to stall with HT off will stall the processor and waste cycles, so with increased contention I would expect the benefit of HT to be higher! |
||
|
|
Mamajuanauk
Master Cruncher United Kingdom Joined: Dec 15, 2012 Post Count: 1900 Status: Offline Project Badges: 
|
@Apis/et al...
----------------------------------------Following my comments earlier, I looked a little closer at the results from 2 machines both with dual 5650's one with HT on, the other with HT off. It should also be noted this is only a rough comparrison as one machine is a Dell server, the other an EVGA_SR2 based system. This comparrison is of the last 30 wu's for each machine, averaged on CPU time and points for each of the results. So, here's the numbers Dell
Average PPR - 128.73 SR2
Average PPR - 87.87 I'm not good on interpreting stats or numbers, but I think I'm correct in the follwoing assumptions: 1 - if the SR2 machine w/o HT is doing 1 wu in 3 hours (clock/calandar/actual hours) then each thread will do 8 wu's per calendar day. Times that by 12 threads you get (3x8)&(24x12)=288 wu's per calendar day 2 - if the Dell machine with HT is doing 1 wu in 5.76 hours (clock/calandar/actual hours) then each thread will do 4.167 wu's per calendar day. Times that by 24 threads you get (5.76x4.167)&(24.00192x24)=576.04608 wu's per calendar day Anyone with a quick eye will notice that 576 (rounded) is actually 2 x 288! EDIT I wouldn't expect the points to reveal much difference... Have I?
Mamajuanauk is the Name! Crunching is the Game!
----------------------------------------  [Edit 2 times, last edit by Mamajuanauk at Oct 16, 2014 9:21:47 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Mamajuanauk,
As you say, interpreting statistics can be fraught with difficulties. I, for one, cannot interpret the points system accurately either, so I'd prefer not to look at those. Comparing two different 'puters is also not a trivial task, but what else to do? All I will say is that, if we assume that the mix of batches in the sample is comparable and that the two computers are also comparable, then the machine with HyperThreading on is taking less than twice the time the other one takes to process a WU. Therefore it is doing more "work" per compute unit(**) per unit of time. But, frankly, is the difference enough to fuss over? I'd say it's about what I'd expect, percentage wise, and I would say stick with HT on. But I won't shout too loud if others disagree. (**) One compute unit is one thread if HT is off, but two threads if HT is on. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Mornin...............shiny new bronze!
 |
||
|
|
Mamajuanauk
Master Cruncher United Kingdom Joined: Dec 15, 2012 Post Count: 1900 Status: Offline Project Badges:
|
Morning All...
----------------------------------------Congrats to all UK members who've reached badges in the over night update!
Mamajuanauk is the Name! Crunching is the Game!
|
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Morning Team!
Missed bronze by 4hrs :( |
||
|
|
Mamajuanauk
Master Cruncher United Kingdom Joined: Dec 15, 2012 Post Count: 1900 Status: Offline Project Badges:
|
Mamajuanauk, As you say let's leave points aside.As you say, interpreting statistics can be fraught with difficulties. I, for one, cannot interpret the points system accurately either, so I'd prefer not to look at those. Comparing two different 'puters is also not a trivial task, but what else to do? All I will say is that, if we assume that the mix of batches in the sample is comparable and that the two computers are also comparable, then the machine with HyperThreading on is taking less than twice the time the other one takes to process a WU. Therefore it is doing more "work" per compute unit(**) per unit of time. But, frankly, is the difference enough to fuss over? I'd say it's about what I'd expect, percentage wise, and I would say stick with HT on. But I won't shout too loud if others disagree. (**) One compute unit is one thread if HT is off, but two threads if HT is on. There may be a difference if one option was producing more results and one option was more runtime, this could focus the preference, if somone was more interested in results than runtime or visa versa Anyway, I think we've established the difference is a matter preference and not so muct fact. It will take a bigger study and someone better at stats than either of us to see if there is any real benefit, which may only be small, if at all. I only have 2 dual machines with HT off, but may well turn them back on in the near future... But for now, enough said... Unless anyone else has any compelling input with evidence
Mamajuanauk is the Name! Crunching is the Game!
|
||
|
|
jonnieb-uk
Ace Cruncher England Joined: Nov 30, 2011 Post Count: 6105 Status: Offline Project Badges: 
|
# 615
----------------------------------------Welcome to the UK Team grangizmo
Don't hesitate to post if you have a query, need information or just to say hello. One of us will always be happy to try and help. |
||
|
|
jonnieb-uk
Ace Cruncher England Joined: Nov 30, 2011 Post Count: 6105 Status: Offline Project Badges:
|
Congratulations to the following UK team members on achieving a Personal Milestone in yesterday's crunching:
----------------------------------------  Mamajuanauk 200,000,000 Points Clockwork 5,000,000 Points Mamajuanauk passes 80,000,000 Points for the UK team seamansgirl 400 Results Clockwork passes 8,000 Results for the UK team Congratulations also to the following UK team members on setting new PBs dcrobinsonRuntime Points 26,149 Results UK team - Individual Ranking Movements
UK team Comparison of Daily RunTime, Points, Results Hours Points Results Average Daily Crunching Comparison RunTime Points Results Milestone Targets for the UK team Target Current To Do 7day Avg. Estimate No. Of Members Active Yesterday Day out of Tot. ***This was a record beating Thu 8 May and Wed 9 Oct when 78 members crunched.  ---------------------------------------- [Edit 1 times, last edit by jonnieb-uk at Oct 17, 2014 9:07:35 AM] |
||
|
|
jonnieb-uk
Ace Cruncher England Joined: Nov 30, 2011 Post Count: 6105 Status: Offline Project Badges:
|
Daily Global5000
----------------------------------------RT(days) # Points # Results # The Daily Global5000 accounted for 71.8% of yesterday's RunTime of 394.6 years The UK Team accounted for 0.54% |
||
|
|
|