Pepo wrote:

In the mean time you can give BoincView a try... In this case, the 1.5.0 series allows you to draw a selection around your 60 tasks and pause them at one button press.

I use BoincView, but this time the computer where it's installed, was unavailable for me from home (intranet specifics at work). I think, such feature will be not unnecessary for Boinc Manager itself too..

I was really surprised to know about single-threaded structure of the main calculation in every BOINC app.

You have to remember when BOINC was first being coded there were no multi-core CPUs and very few hyperthreaded or multi-CPU systems, so there was little reason to do anything other than single thread apps. In fact the reason support for multi-threaded apps are being talked about now is the advent of systems with more than 4 CPUs. Up to 4 CPUs running one app per CPU works well, higher than that memory contention starts becoming a limiting factor.

Another thing to consider is that after a project splits it's workload up enough to be sent out to your computers there is very little parallelism left in the app. From another point of view each computer gets one thread of a massively multi-threaded app. Splitting the work enough to send it out to individal computers may not leave any way to split it up more for most projects. The only project I can think of that may be able to split up their app more is CPDN. The may be able to create a "land", "ocean", and "air" thread within their app.

---

BOINC WIKI

BOINCing since 2002/12/8

The only project I can think of that may be able to split up their app more is CPDN. The may be able to create a "land", "ocean", and "air" thread within their app.

Hi John

I know a lot of people would like CPDN to split the workunits into smaller parts, either horizontally so the time period is shorter, or as you suggested vertically so that fewer parameters are calculated and the full time period processes faster.

As far as I know the vertical split cannot be done because most if not all the parameters interact with at least one other. For a vertical split there would have to be at least two groups of parameters that don't interact through time. But the HADCM models are called 'coupled' because they take into account that ocean conditions affect the climate everywhere including over land.

The horizontal time split has been implemented with HADCM 160-year models. The current 80-year HADCMs take 160-year models processed for at least 80 years by another cruncher and crashed, then let another computer continue them from the 80-year restart dump. It could be that the last 40 years of 160-year models that crashed near the end are later rereleased to be completed on another machine. The same thing could I believe also be done for shorter final periods of long crashed models, not even taking them from a restart dump (more likely I think from a decadal file upload).

@mo.v
I may be wrong but I think the problem with splitting the models vertically is the amount of data transfer between the two parts. If the parts are on the same computer this would be less of an issue.

CPDN is the only project I have seen developers expressing an interest in creating a multi-threaded app for. Of course this could be strictly for new studies without affecting current apps. In any case it will not be happening anytime soon.

---

BOINC WIKI

BOINCing since 2002/12/8

If anything can be done in future to have these massive models sitting on computers for less time in total, this will improve the completion rate.

AFAIK there are 500-year versions of the same or similar HADCM models but they've only been run on a supercomputer.