+- +-
Say hello if visiting :) by Gecko
11 Jan 2023, 07:43:05 pm

Seti is down again by Mike
09 Aug 2017, 10:02:44 am

Some considerations regarding OpenCL MultiBeam app tuning from algorithm view by Raistmer
11 Dec 2016, 06:30:56 am

Loading APU to the limit: performance considerations by Mike
05 Nov 2016, 06:49:26 am

Better sleep on Windows - new round by Raistmer
26 Aug 2016, 02:02:31 pm

Author Topic: Installer v0.44 release notes - required update for v8  (Read 12990 times)

Offline Richard Haselgrove

  • Messenger Pigeon
  • Knight who says 'Ni!'
  • *****
  • Posts: 2819
Installer v0.44 release notes - required update for v8
« on: 22 Jan 2016, 05:51:35 am »
Lunatics are pleased to announce the availability of the v0.44 installer in the Lunatics main download area and mirror sites.
Please take care you use the right installer (32/64 bit) for your system.

Should you have any questions or problems - ask! Post in this thread or open a new one.

New in this installer:
---------------------
Support for setiathome_v8 tasks recorded at the Arecibo radio telescope, and (in conjunction with Breakthrough Listen) tasks recorded at the Green Bank Observatory and any other suitable radio telescope.

Astropulse applications are unchanged from release v0.43b.

How to use the installer:
------------------------
For the CPU applications, you need to know what your CPU supports. Either check the Boinc startup messages or check with CPU-Z downloadable from http://www.cpuid.com
Tick or untick CPU AP and MB applications as desired. 
Tick ATI (OpenCL) AP or MB if/as desired.
Tick NVidia (OpenCL) AP or CUDA MB if/as desired.
Tick Intel (OpenCL) AP or MB if/as desired.
More details further down.

NB We use default values throughout the app_info.xml. If you have previously installed optimised applications and manually edited app_info.xml (e.g. <count> for multithreading GPUs, <flops>, <cmdline> for ATI) you WILL lose those edits on running the installer and have to redo them. You may wish to make a note of them.
You will also lose all other manual edits to app_info.xml such as beta applications.

You are strongly recommended to use a version of BOINC which supports app_config.xml files. These can be used to manage the number of tasks running on each GPU more simply than by editing app_info.xml, and will NOT be removed by re-running this or subsequent installers. If you have previously used an app_config file for MultiBeam v7, you may want to add a similar section for setiathome_v8.

Known issues - general:
----------------------
Be careful when upgrading if you have already pre-installed optimised MB v8 applications from individual pre-release build packs. Some of these packs deploy version numbers derived from Beta testing, and they are not guaranteed to match the Main project release values. This installer follows the application version numbers and plan classes displayed on http://setiathome.berkeley.edu/apps.php - if yours differ, set NNT and drain your cache before proceeding.

TAKE SPECIAL CARE when upgrading from the stock 7.00 Astropulse CPU application. The intermediate checkpoint files have a non-standard format, and the Lunatics applications can't resume a part-computed task. Either set NNT and wait until the current task(s) have finished - may be a long wait - or abort tasks if they haven't progressed very far. This only applies to v7.00 tasks which have commenced processing - v7.03 (sse or sse2) tasks are safe to upgrade while in progress, and all unstarted tasks are safe.

BOINC v7 installations are not able to run any GPU applications in service mode, and the installer will now bypass GPU application pages on machines which cannot run GPU applications because of this restriction.

The installer is supposed to shut down the BOINC client whilst leaving the Manager running - noticeable from the red 'disconnected' dot in the tray. After finishing the BOINC client should restart and the red dot vanish when the Manager reconnects. This may take a few minutes. On some systems however this fails to work. Please ensure BOINC has restarted.
Advanced users only: If you know you are going to edit app_info.xml, it may be easier to shutdown and restart BOINC manually, so you can get the changes into app_info.xml before the restart.

Some users reported that installer v0.41 failed to assemble a complete app_info.xml file from the aistub files supplied with each application. We have switched to using a 'grep' application (supplied) called from a slightly modified 'aimergeg.cmd' script to help avoid this problem. The original aimerge script is also present for people who are used to using it.


CPU MB apps - MB8_win_xxx_VS r3330

In the continuing absence of Joe Segur (we wish him well, wherever he is), we have only been able to provide a limited range of CPU apps - SSE (not applicable to 64-bit platforms), SSE2, SSE3, and AVX. These have been compiled using Microsoft Visual Studio, and may not be as tightly optimised as Joe's wizardry with GCC 4.7.1. When one of the intermediate SIMD levels like SSSE3 or SSE4.1 is detected by the installer and selected, the SSE3 application will be supplied for v8 tasks, but the matching v7 legacy (AKv8c r2549) application will be supplied for cleaning-up leftover v7 tasks.

If you have any doubt at all about what instruction sets your CPU supports, download CPU-Z from http://www.cpuid.com/softwares/cpu-z.html

We have now incorporated a 3rd party CPU feature detection tool into the installer. Unsupported applications will be greyed out, and the installer will pre-select recommended application versions for your CPU.

Users of AMD CPUs capable of running the AVX instruction set may wish to check whether the SSE3 or AVX application is faster on their particular CPU.


CPU AP v7 app - r2672/2692/2696/2703 depending on optimisation level.

A 32-bit AVX build is included as previously. At the SSE3 level, the installer will choose between AMD and Intel builds automatically (there wasn't space for an extra manual choice).


NVidia GPU (CUDA) app - x41zi

The CUDA application is capable of handling both the new v8 and the legacy v7 tasks - there is no need for a separate v7 application. The application will switch modes automatically with no need for user configuration.

For details please see the separate ReadMe_x41zi.txt.

Rule of thumb selection guide:
Cuda 2.3: PreFermi - NOT to be used on Fermi/Kepler/Maxwell
Cuda 3.2: mixed systems
Cuda 4.2: Fermi
Cuda 5.0: Kepler/Maxwell
If you have run stock and have established APR you can also pick the fastest app from there.
For compatible cards see http://developer.nvidia.com/cuda-gpus
Should run on all NVidia CUDA cards with at least 512MiB video RAM. Actual free memory required is in the region of 200-235MB, depending on driver, OS flavour and CUDA variant. The code will scale up if more memory is available. Running on a 256MiB card is just possible, but you will have to free up as much VRAM as possible, e.g. by disabling Aero.
Used video RAM can be checked with GPU-Z downloadable from
http://www.techpowerup.com/downloads/2571/techpowerup-gpu-z-v0-8-6
Watch closely on GPU-Z and in BOINC manager if the application finds enough memory.
Symptoms of insufficient memory are
- in BOINC manager: tasks start up and run a few seconds then go to 'waiting to run' and the next task tries to start
- on GPU-Z: the sensor for memory shows used VRAM ramping up and almost immediately dropping again.

Depending on BOINC version your tasks will either error out with 'too many exits' or get stuck in an infinite loop.
If you don't have enough memory, set BOINC to NNT, abort tasks and uninstall the CUDA application by rerunning the installer with CUDA MB UNticked. If in doubt ask for help on NC: http://setiathome.berkeley.edu/forum_forum.php?id=10
To multithread on Fermi/Kepler cards, find all instances of <count>1</count> in app_info.xml and decrease to 0.5 or 0.3 to run two or three tasks in parallel.
Again, if  you are uncertain of how to do this correctly, ask on the above mentioned forum. 'Your mileage may vary' - it depends on your specific system which count will give the highest throughput.
Full release notes and history in separate readme.


NVidia (OpenCL) AP v7 app - retained at r2887

Please see separate ReadMe_AstroPulse_OpenCL_NV.txt for details.
Reserving one CPU core per GPU card will improve performance.

Warning - the NVidia 340.52 driver (released 29 July 2014) causes invalid results on legacy compute capability 1.1, 1.2 and 1.3 cards. This application will - deliberately - NOT run on such cards: it will alternate between displaying a warning message in BOINC Manager and apparently 'running', but will make no progress and eventually error out. If you have an old card (8000 series, 9000 series, or 2xx series), either exclude it from crunching AP, or downgrade to an older driver.

Alternatively, the 'hotfix' 341.44 driver (released 24 February 2015) can be used if available - not all operating system versions are supported.


ATI (OpenCL) MB app - updated to r3330

The new r3330 applications will run v8 tasks ONLY.

The previous r2929 application is supplied again to handle legacy v7 tasks ONLY.

Please see separate ReadMe_MultiBeam_OpenCL_ATI.txt for details, especially on driver requirements.
Reserving one CPU core per GPU card will improve performance.
We offer the MB application in two different variants, depending on card:
plain - for most HD4xxx owners and those with HD5xxx, HD6xxx cards and HD7xxx cards who suffer from driver incompatibilities.
HD5 - for most if not all owners of HD5xxx, HD6xxx and HD7xxx GPUs.


ATI (OpenCL) AP v7 app - retained at r2742
ATI (Brook) AP app - retained at r2904

Please see separate ReadMe_AstroPulse_OpenCL_ATI.txt or ReadMe_AstroPulse_Brook.txt for details, especially on driver requirements.
Reserving one CPU core per GPU card will improve performance.

For cards without OpenCL support (from HD2xxxx on) use the 'hybrid' r2904 Astropulse application. NB this is a combined CPU/GPU application and will only use the GPU for some of the calculations, while doing the rest on a CPU core. Minimum driver version Catalyst 9.x.

The Hybrid AP application uses about 80% of a CPU core and does only about 20% of the calculations on the GPU. Parameters have been choosen to reflect this usage. However present BOINC can not reserve '0.8' of a CPU core and therefore reserves none. This means your CPUs will be overcommitted and runtimes and DCF (if applicable) will suffer. You do increase the overall output though.

HD2xxx cards requiring Brook (Hybrid) applications are once again supported by Lunatics for Astropulse v7.


Intel (OpenCL) applications for both MB and AP v7 (updated to r3330/r2742)

The new r3330 applications will run v8 tasks ONLY.

The previous r2929 application is supplied again to handle legacy v7 tasks ONLY.

These are now offered by Lunatics. Please note the supported iGPU models, and observe the minimum OpenCL and driver versions, shown on the selection page. Intel CPUs offering plain "HD graphics" may be able to run these applications, with a suitable driver, but have not been exhaustively tested. Please verify that tasks run on these devices are validating consistently before allowing them to run unattended.

You MUST update your Intel driver if BOINC reports the OpenCL version as 1.1
The applications may appear to run without updating the driver, but the results will not be valid and you will have wasted your computing power.

Visit https://downloadcenter.intel.com/ if you require an updated graphics driver for your iGPU. Note that you will require a driver for a 3rd. or 4th. generation Intel 'core' CPU with the appropriately numbered HD or Iris graphics support.

Many people have reported problems getting Intel's automated update and installation tools to work, but their manual installation instructions (requiring the download of the '.zip' driver package) appear to work on systems which reject the automated tools. A copy of Intel's ReadMe file for driver installation will be placed in the Lunatics docs folder for reference if you select either of the iGPU applications.


This time the thanks go to Raistmer for the optimised CPU apps, to Raistmer (with additional input from Urs Echternacht) for OpenCL apps and to Jason (jason_gee) for CUDA apps.
Thanks guys. There would be no need for an optimised app installer if there weren't optimised apps in the first place!

We would also like to take this opportunity to thank our active alpha testers
(in no particular order and apologies to anybody we've forgotten):
Claggy, arkayn and Mike.


The Lunatics Installer team AD MMXVI

 

Welcome, Guest.
Please login or register.
 
 
 
Forgot your password?
Members
Total Members: 97
Latest: ToeBee
New This Month: 0
New This Week: 0
New Today: 0
Stats
Total Posts: 59559
Total Topics: 1672
Most Online Today: 24
Most Online Ever: 983
(20 Jan 2020, 03:17:55 pm)
Users Online
Members: 0
Guests: 34
Total: 34
Powered by EzPortal