Brad,<div><br></div><div>I wont have time to try this theory till tomorrow, but i just realized that no matter what you do with the threading of the Jacobian, the operator= is still being called in the translation transform. This may be the culprit. It might be clearing and reallocating no matter what you do. Might explain some of the slowdown single threaded as well.</div>
<div><br></div><div>Changing transform, or modifying translation transform might show this up.</div><div><br></div><div>Just a thought.</div><div>Rupert<br clear="all">--------------------------------------------------------------<br>
Rupert Brooks<br><a href="mailto:rupert.brooks@gmail.com">rupert.brooks@gmail.com</a><br><br>
<br><br><div class="gmail_quote">On Thu, Jul 26, 2012 at 12:48 PM, Rupert Brooks <span dir="ltr"><<a href="mailto:rupert.brooks@gmail.com" target="_blank">rupert.brooks@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Brad,<div><br></div><div>Am i reading this right? - the second set of numbers is vanilla ITK 4.2 and the third set is with your patch?</div><div><br></div><div>a) Single threaded, allocating outside the threaded part made the mean time go up from .38 to .4 in the single threaded case, and it seems to be a real effect. </div>
<div>b) After allocation outside the thread, the multithread performance is worse both absolutely, and in the serial/parallel percentage.</div><div><br></div><div>Wierd.</div><div><br></div><div>I also note that going from 4 to 8 threads seems to get you a greater ratio of improvement than going from 1 to 2, or 2 to 4. Especially with the vanilla itk 4.2. Generally parallel improvement drops off with number of threads, because you are only subdividing the parallel part of the algorithm. I have found in the past that this sort of wierdness can occur with NUMA systems depending whether the OS puts multiple threads on the same physical chip or not.</div>
<div><br></div><div>Are you really fortunate enough to have 288 cores in that Mac, or is it more like 3 chips at 8 cores each or something?</div><div><br></div><div>My personal tendency would be to try to understand why the single thread performance got so much worse, then (if relevant) try and fix the threading. ITK 3.20 doesnt do very well with increasing the numbers of threads either.</div>
<div><br></div><div>Just to reiterate where im coming from - i want to show that accuracy and performance wise ITK4 is at least no worse than ITK3. That justifies upgrading. In the longer run, it may make more sense to rewrite old ITK stuff to use the new metrics, but that is a separate project. </div>
<div><br></div><div>Rupert</div><div>--------------------------------------------------------------<br>Rupert Brooks<br><a href="mailto:rupert.brooks@gmail.com" target="_blank">rupert.brooks@gmail.com</a><div><div class="h5">
<br><br>
<br><br><div class="gmail_quote">On Thu, Jul 26, 2012 at 11:53 AM, Bradley Lowekamp <span dir="ltr"><<a href="mailto:blowekamp@mail.nih.gov" target="_blank">blowekamp@mail.nih.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div style="word-wrap:break-word">Hello,<div><br></div><div>Well I did get to it before you:</div><div><br></div><div><a href="http://review.source.kitware.com/#/c/6614/" target="_blank">http://review.source.kitware.com/#/c/6614/</a></div>
<div><br></div><div>I also uped the size of the image 100x in your test, here is the current performance on my system:</div><div><br></div><div><div>System: <a href="http://victoria.nlm.nih.gov" target="_blank">victoria.nlm.nih.gov</a></div>
<div>Processor: Intel(R) Xeon(R) CPU X5670 @ 2.93GHz</div><div> Serial #: </div><div> Cache: 32768</div><div> Clock: 2794.27</div><div> Cores: 12 cpus x 24 Cores = 288</div><div>OSName: Mac OS X</div>
<div> Release: 10.6.8</div><div> Version: 10K549</div><div><div> Platform: x86_64</div><div> Operating System is 64 bit</div><div>ITK Version: 3.20.1</div></div><div>Virtual Memory: Total: 256 Available: 228</div>
<div>Physical Memory: Total:65536 Available: 58374</div><div><div> Probe Name: Count Min Mean Stdev Max Total </div></div><div> MeanSquares_1_threads 20 0.344348 0.347567 0.00244733 0.352629 6.95134</div>
<div> MeanSquares_2_threads 20 0.251223 0.300869 0.0179305 0.321404 6.01738</div><div> MeanSquares_4_threads 20 0.215516 0.348677 0.173645 0.678274 6.97355</div>
<div> MeanSquares_8_threads 20 0.138184 0.182681 0.0297812 0.237129 3.65362</div><div>System: <a href="http://victoria.nlm.nih.gov" target="_blank">victoria.nlm.nih.gov</a></div><div>Processor: </div>
<div> Serial #: </div><div> Cache: 32768</div><div> Clock: 2930</div><div> Cores: 12 cpus x 24 Cores = 288</div><div>OSName: Mac OS X</div><div> Release: 10.6.8</div><div> Version: 10K549</div><div>
<div> Platform: x86_64</div><div> Operating System is 64 bit</div></div><div>ITK Version: 4.2.0</div><div>Virtual Memory: Total: 256 Available: 228</div><div>Physical Memory: Total:65536 Available: 58371</div><div>
<div> Probe Name: Count Min Mean Stdev Max Total </div></div><div> MeanSquares_1_threads 20 0.382481 0.383342 0.00186954 0.391027 7.66685</div>
<div> MeanSquares_2_threads 20 0.211908 0.335328 0.0777408 0.435574 6.70655</div><div> MeanSquares_4_threads 20 0.271531 0.315688 0.0390751 0.385683 6.31377</div>
<div> MeanSquares_8_threads 20 0.147544 0.192132 0.0299427 0.240976 3.84263</div></div><div><br></div><div><br></div><div>In the patch provided, it is implicitly done on assignment on a per-thread basis. What was most un-expected was when then allocation of the Jacobin was explicitly done out side the threaded part, the time when up by 50%! I presume that the sequential allocation, of the doubles in the master thread made the allocation sequentially, next to each other, and may be a more insidious form of false sharing. Below is the numbers from this run, notice the lack of speed up with more threads:</div>
<div><br></div><div><div>System: <a href="http://victoria.nlm.nih.gov" target="_blank">victoria.nlm.nih.gov</a></div><div>Processor: </div><div> Serial #: </div><div> Cache: 32768</div><div> Clock: 2930</div><div> Cores: 12 cpus x 24 Cores = 288</div>
<div>OSName: Mac OS X</div><div> Release: 10.6.8</div><div> Version: 10K549</div><div><div> Platform: x86_64</div><div> Operating System is 64 bit</div></div><div>ITK Version: 4.2.0</div><div>Virtual Memory: Total: 256 Available: 226</div>
<div>Physical Memory: Total:65536 Available: 57091</div><div><div> Probe Name: Count Min Mean Stdev Max Total </div></div><div> MeanSquares_1_threads 20 0.403931 0.40648 0.00213043 0.41389 8.1296</div>
<div> MeanSquares_2_threads 20 0.243789 0.367603 0.0894637 0.65006 7.35206</div><div> MeanSquares_4_threads 20 0.281336 0.354749 0.0431082 0.440161 7.09497</div>
<div> MeanSquares_8_threads 20 0.24615 0.301576 0.0552998 0.446528 6.03151</div></div><div><br></div><div><br></div><div>Brad</div><div><br></div><div><div><div><br><div><div>
On Jul 26, 2012, at 8:56 AM, Rupert Brooks wrote:</div><br><blockquote type="cite">Brad,<div><br></div><div>The false sharing issue is a good point - however, i dont think this is the cause of the performance degradation. This part of the class (m_Threader, etc) has not changed since 3.20. (I used the optimized metrics in my 3.20 builds, so its in Review/itkOptMeanSquares....) It also does not explain the performance drop in single threaded mode.</div>
<div><br></div><div>Testing will tell... Seems like a Friday afternoon project to me, unless someone else gets there first.</div><div><br></div><div>Rupert</div><div><br clear="all">--------------------------------------------------------------<br>
Rupert Brooks<br><a href="mailto:rupert.brooks@gmail.com" target="_blank">rupert.brooks@gmail.com</a><br><br>
<br><br><div class="gmail_quote">On Wed, Jul 25, 2012 at 5:18 PM, Bradley Lowekamp <span dir="ltr"><<a href="mailto:blowekamp@mail.nih.gov" target="_blank">blowekamp@mail.nih.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div style="word-wrap:break-word">Hello,<div><br></div><div>Continuing to glance at the class.... I also see the following member variables for the MeanSquares class:</div><div><br></div><div><div> MeasureType * m_ThreaderMSE;</div>
<div> DerivativeType *m_ThreaderMSEDerivatives;</div></div><div><br></div><div>Where these are index by the thread ID and access simultaneously across the threads causes the potential for False Sharing, which can be a MAJOR problem with threaded algorithms.</div>
<div><br></div><div>I would think a good solution would be to create a per-thread data structure consisting of the Jacobin, MeasureType, and DerivativeType, plus padding to prevent false sharing, or equivalently assigning max data alignment to the structure.</div>
<div><br></div><div>Rupert, Would like to take a stab at this fix?</div><div><br></div><div>Brad</div><div><div><div><br></div><div><br><div><div>On Jul 25, 2012, at 4:31 PM, Rupert Brooks wrote:</div><br><blockquote type="cite">
Sorry if this repeats - i just got a bounce from Insight Developers, so im trimming the message and resending....<br clear="all">--------------------------------------------------------------<br>Rupert Brooks<br><a href="mailto:rupert.brooks@gmail.com" target="_blank">rupert.brooks@gmail.com</a><br>
<br>
<br><br><div class="gmail_quote">On Wed, Jul 25, 2012 at 4:12 PM, Rupert Brooks <span dir="ltr"><<a href="mailto:rupert.brooks@gmail.com" target="_blank">rupert.brooks@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Aha. Heres around line 183 of itkTranslationTransform.<div><div><br></div><div>// Compute the Jacobian in one position</div><div>template <class TScalarType, unsigned int NDimensions></div><div>void</div><div>TranslationTransform<TScalarType, NDimensions>::ComputeJacobianWithRespectToParameters(</div>
<div> const InputPointType &,</div><div> JacobianType & jacobian) const</div><div>{</div><div> // the Jacobian is constant for this transform, and it has already been</div><div> // initialized in the constructor, so we just need to return it here.</div>
<div> jacobian = this->m_IdentityJacobian;</div><div> return;</div><div>}</div><div><br></div><div>Thats probably the culprit, although the root cause may be the reallocating of the jacobian every time through the loop.</div>
<div>
<div><br></div><div>Rupert</div><div><br></div><div><snipped></div></div></div></blockquote></div>
</blockquote></div><br></div></div></div></div></blockquote></div><br></div>
</blockquote></div><br></div></div><div><div>
<div style="word-wrap:break-word;font-size:12px"><p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px"><font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica">========================================================</font></p>
<p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px"><font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica">Bradley Lowekamp<span> </span><span> </span></font></p>
<p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px">
<font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica">Medical Science and Computing for</font></p><p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px"><font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica">Office of High Performance Computing and Communications</font></p>
<p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px"><font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica">National Library of Medicine<span> </span></font></p><p style="margin-top:0px;margin-right:0px;margin-bottom:0px;margin-left:0px">
<font face="Helvetica" size="3" style="font:normal normal normal 12px/normal Helvetica"><a href="mailto:blowekamp@mail.nih.gov" target="_blank">blowekamp@mail.nih.gov</a></font></p><br></div><br>
</div>
<br></div></div></div></blockquote></div><br></div></div></div>
</blockquote></div><br></div>