Re: Z840 Optimal BIOS Settings - Help Please

Error

Long processing time with Xeon v3 processors.

Description

When processing projects with Xeon v3 processors such as, but not limited to, the Intel Xeon CPU E5-2630 v3 or the Intel Xeon CPU E5-2687W v3, processing times are significantly longer than expected.

Cause

Under investigation.

Workaround

The following steps can alleviate the problem:

- Update motherboard BIOS

- Update chipset driver

- Disable hyperthreading

- Update GPU driver (crucial)

Until the cause of the issue is identified, it is recommended to choose a different CPU if possible (due to longer processing times).

_________________________________<END

As you have disabled hyperthreading, you may well have seen this already.

My sense is that the problem may be poor application scalar multi-threading. The intial steps of the data to map conversions as in all 3D modeling is going to be highly single-threaded, but this task seems to variably distributed by the application and of course, the operating system.  When I run various 3D modeling applications on a 6-core / 12-thread Xeon E5, I can see activity in 6 threads,and the distribution may well be Windows.

Ordinarily, in this situation,  the problem might be attributed to the single-thread performance- so important in 3D modelling ,of the Xeon E5-2620 v3 (8-core @ 2.4 /3.2Ghz), which has a Passmark Single Thread Mark of 1670 but the notice also mentions the E5-2687W v3 10-core @ 3.1/3.5 with a rating of 1940.  I have an HP z429 with an E5-1620 (4-core @ 3.6 /3.8) with a single thread score of 1931 and that is adequate in AutoCad, Revit, Sketchup, and Adobe CS6.  I'm not aware of the details of the E5- v3 architecture,  but given the circumstances, Intel thoroughness, and the level of HP development I would supect it's an application situation.

However, there may be a work around:

In the notes on the system requirements page for Pix4D Mapper Pro:

https://support.pix4d.com/hc/en-us/articles/202557289#gsc.tab=0

_______________________________________________

The interesting paper in this regard is 204191535 which includes:

______________________________________________

Hardware used

The software is highly parallelized and takes advantages of multi-core CPUs, as well as SSE/AVX instructions and NVIDIA GPU CUDA processing.

However the different steps of processing do not use these resources the same way, and not all parts can take advantages of multi-core or CUDA.

Note that to increase the performance it is important to have a balanced configuration without bottlenecks.

CPU

Most of the processing is done on the CPU, and a faster CPU is the first key to increase processing speed. Hexa and octo cores, latest generation i7 or Xeon CPUs are recommend . Clock speed generally impacts the full project, and number of cores greatly impacts step 2.Point Cloud and Mesh.

Dual socket CPU does not double performance, but does generally provide faster processing than single socket CPU.

RAM

The amount of RAM mostly has an impact on the number of images that can be processed in a single project. It also has an impact on the processing of step 2.Point Cloud and Mesh.

Hard Disk

Step 3. DSM, Orthomosaic and Index is the most impacted by the hard disk speed. Using octo core CPUs and a fast SSD improves the performance of step 3.

Graphic Card

Starting with version 1.3, Pix4Dmapper takes advantages of NVIDIA GPUs with CUDA to further increase processing speed.

The speed increase highly depends on the project type, i.e. the image number, the image size and the image content. For example, with a GTX 970 we observed a speedup between 10% and 75% for step 1. Initial Processing, and around 5% for step 2.Point Cloud and Mesh without the SGM option. As a rule of thumb, projects with high overlap, high image content and thus high number of keypoints benefit more from the speedup.

_______________________________________________<END

This confirms the idea of the need in this kind of software where very high polygon meshes are generated  that the clock speed- and therefore the single-thread perofrmance is higly important.   Hwoever, the navigation compute and  visualization can be GPU-accelerated

Given your current commitments in costs, and this section of the above performance notes:

______________________________________________

Graphic Card

Starting with version 1.3, Pix4Dmapper takes advantages of NVIDIA GPUs with CUDA to further increase processing speed.

The speed increase highly depends on the project type, i.e. the image number, the image size and the image content. For example, with a GTX 970 we observed a speedup between 10% and 75% for step 1. Initial Processing, and around 5% for step 2.Point Cloud and Mesh without the SGM option. As a rule of thumb, projects with high overlap, high image content and thus high number of keypoints benefit more from the speedup.

______________________________________________ >END

 The recommendation quoted earlier, "Until the cause of the issue is identified, it is recommended to choose a different CPU if possible (due to longer processing times)" is to me quite casually suggesting shelving and replacing expensive hardware ( a pair of Xeon E5-2687w v3 costs more than $4,400) to mitigate with a particular problem in one program.   and does not inspire confidence.  However, instead of changing processors, my suggestion is to consider trying an NVIDIA Maximus configuration which would pair the Quadro K5200 with a Tesla GPU coprocessor. this will add the GPU ,CUDA, and several GB's of graphics RAM in accleration intitally, buy a used Tesla C2075 which has 6GB of memory.

A poster on this site, Brian1965 has been working with a Maxumus configuration of a Quadro K4200 and Tesla M2090 in a z620.  In OctaneBench Results , there are three such configuations tested withe an average score of 90.12 in which a score of 100 is the averge performnace of a GTX 980.  Brian1965 points out that that is score is better than a Quadro K6000 (average is 87) and of course a K5200 is going to improve on that. That is impressive enough, but the 6GB of memory in the Tesla iseems to me a very important consideration in your use given the very large datasets in mech generation. Keep in mind that Teslas are made in two forms, one for servers with passive cooling such as the M2090 requiring custom cooling solutions and one for workstations with active cooling such as K20, C2075, and etc. 

These do require two PCIe slots, 225W and good cooling conditions,  but this is the kind of hardware the z840 should incorporate easily. With all the variables- epecially the program maker's mention of the mysterious low peformance of the E5-2630 v3- there are no guarantees,  but the experiment might be done for a couple of hundred dollars that could be recovered if unsuccesful or recovered and applied to a K20 or other more modern verisons. 

If that approach does not accelerate the work, and given that Pix4D Mapper Pro finds Xeon E5-v3 problematic, the next step might be to consider changing to Xeon E5-2600 v4  such as the E5-2637 v4 6-core @ 3.5/3.7GHz, having a Passmark single thread rating of 1883. It seems radical,  but is 1/4 the cost of changing the software.

The optimization of workstations today is quite difficult as using 4 or 5 programs is common and some will need high single- thread rates, others benefit from many CPU cores and still others from GPU acceleration.

I would be interested to know what happens.

Cheers,

BambiBoomZ