World Community Grid - View Thread

It is a known problem with ARP1, and also a reason why by default, the number of concurrent ARP1 WUs has been restricted. But there are always a handful of rather selfish folks that have to mess with those settings, making things ultimately worse for everyone... sad

Ralf

You are wrong on both points!

First: this limit (by default, only 1 task per computer) for the ARP1 sub-project was introduced many years ago not because of congestions in the network infrastructure (when it was initially introduced, they simply did not exist, because then the project was based on a much more powerful network infrastructure provided by IBM, which without significant problems cope with such a load).
And it was actually introduced because of the high system requirements (at least several times higher than those of all other WCG sub-projects) for clients: in order to limit the load on the volunteer's computers and reduce interference/impact on their work (do not forget that most of the computers involved in distributed computing are not dedicated crunchers working ONLY for DC exclusively, but are used by users regularly for their other tasks, often DC run in background "set-and-forget" mannener without any monitoring and tuning).
In terms of RAM and disk usage, as well as traffic (relevant for those using limited Internet connections).

Although by current/modern standards 700-900 MB of RAM used per 1 ARP1 WU is not a really big deal, but again - this limit was introduced many years ago, when 700-900 MB of RAM usage for each CPU core engaged in processing ARP tasks - it was a LOT and strongly influenced the work of ordinary computers of that time.
A reminder of this is still the warning link next to ARP1 in the WCG settings for selecting sub-projects to participate. These links explain why this limit was introduced/needed:
https://www.worldcommunitygrid.org/help/topic.s?shortName=minimumreq
https://www.worldcommunitygrid.org/help/topic.s?shortName=results#843

Second: an increase in the number of tasks processed by one computer by part of the volunteers does NOT make the situation with "network congestion" worse for others.
Because the total number of tasks processed does not change at all - this is controlled by the WCG developers on the server side: they set how many tasks in total and what batches / at what pace to generate and distribute to clients for the calculation.

This client side setting only affects how many client computers this fixed (limited on the server side) number of tasks will be distributed to. And here, the less, even better, not worse. For two main reasons:

1 - the main problems for Krembil are the number of simultaneous (parallel) network connections, although there are also problems with total bandwidth too, but the most problematic part is the number of parallel connections/requests to front server.
Now consider a situation where, say, we have a total of 10K (10 thousand) tasks prepared by the project servers for distribution in one batch. With default settings (only 1 task per computer) used by ALL participants, these 10K tasks will be distributed to 10K different computers, which in the process of downloading input files for them can create up to 20K parallel network connections to the project servers. Because with the default settings of BOINC, the client opens 2 network connections for 1 project for downloads (although this behavior can also be changed by editing the client configs, but this is another topic).
Now let's say we have 5 (on average) ARP1 WUs per computer, then 10K of the original WUs from one batch will be distributed to only 2K computers that can open up to a maximum of 4K network connections to WCG servers. Yes, they will keep them for much longer time (because you need to download more input files for more WUs), but it will be sequentially, one after the other, and not all in parallel. The number of parallel/simultaneous network connections is reduced significantly (~ 5 times in this conditional example).

2 - The total number of files (and their total volume/bandwidth needed) that need to be downloaded from servers for a computer processing many tasks is slightly less than for many computers processing only 1 task each. Because among the input (source) files there are not only unique (for each specific WUs) files, but also part of the files used in a set of different WUs. When processing few of such WUs on the same computer, these files will be downloaded only once. When processing 1 task on different computers, each computer will download its own separate copy of such "shared" files, creating an extra (which can be avoided) network load on WCG servers.