It's somewhat interesting - this is the first time I've felt like any or all my machines need a RAM upgrade. I have an i7 4770k with 8GB of RAM and a 2070 which has been running other projects with no issues, but then Amicable comes along and gives a computation error on startup. Rosetta running out of memory on a couple other machines as well, etc. Amazon will have more RAM here tomorrow, but I definitely wasn't expecting RAM to be a big bottleneck going into this.
The Amicable 10^21 application originally used 8 GB RAM + 8 GB VRAM, then was soon rewritten to use 8 GB RAM + less than 2 GB VRAM. According to message 1204, the lower VRAM usage is achieved by performing the work in 8 stages. I wonder if it is feasible to have much of the memory of a task swapped out to a fast disk much of the time.I am running a 8GB zram swap, which varies from being half full to completely full depending on the luck of the draw on Rosetta processes. I can increase the size to 16GB, but I’ve found Rosetta only reliably gets 50% compression, so that would only open up 4GB extra.
@mopardude87
You will find that CPU thread count, RAM and powerful GPUs are king in BOINC. When it comes to BOINC competitions that is...............Oh and it is for some.............addicting.
In 2019, the Pentathlon had NFS, Yoyo, and LHC. While all of those have projects with little RAM requirement, the ones with good credit have larger RAM footprint too. Amicable's new application is out of this world though, even compared to LHC's vbox based Atlas application.It's somewhat interesting - this is the first time I've felt like any or all my machines need a RAM upgrade. I have an i7 4770k with 8GB of RAM and a 2070 which has been running other projects with no issues, but then Amicable comes along and gives a computation error on startup. Rosetta running out of memory on a couple other machines as well, etc. Amazon will have more RAM here tomorrow, but I definitely wasn't expecting RAM to be a big bottleneck going into this.
RAM requirement by Rosetta@home changes over time.
(It was generally lower than now before they started COVID work. Then it suddenly was 3 GB/task, if I recall correctly, or something like that. And now it is as discussed here. It won't remain as-is forever, but my guess is that the tendency is more likely towards bigger than towards smaller.)
Alright then so its settled i will opt for the 64gb, i rather have it and not need it over needing it and not having it.The only kit on Newegg that is on the QVL list for my motherboard
Are you a human?
www.newegg.com
Ya know, that is actually not a bad price for 3600 at that capacity....sure wish I had seen that 2 weeks ago instead of price shopping....
A few thoughts / notesAll for the sake of dropping in the 4000 top end beast at the end of the year.
Hopefully we get some 24 core action, i got a feeling we will see 24 cores. If it costs like $900 or less i will prob hop on it. Would cost essentially twice as much as my 12 core 3900x and to me that would be a fair price. Anything less and its win. I could be wrong and 4000 just improves on what we currently got. I got a strong feeling 4000 is dropping more then 16 cores and i want to be ready for it.
A real upgrade at bare minimum to me would be a 32 thread "4950x" that clocks at least to 4ghz on all threads consistently. Anything else improved is win, anything less and well it better be cheaper or way more efficient.
A few thoughts / notes
Ram:
There are some benefits to running 4 sticks of ram instead of 2 sticks of ram (fully populating the board). The really high level is that going with 4 sticks means you're less likely to be able to overclock the ram / hit the highest speeds, but you'll see better throughput at stock speed. See this article, which explains it way better than I can: https://www.techspot.com/article/1971-more-ram-modules-better-for-gaming/. That said, if you buy 2 Sticks now, you can always buy another 2 sticks for an upgrade later and not have any left over dimms.
2x 32GB shouldn't be a problem at all, especially at the price you found and double(?) especially that its on your QVL list. I probably complicated the situation by giving you yet another option . 64GB on a 12 core 24 thread processor gives you 2.6GB per thread. On my 2700X (8 core, 16 thread, currently 16GB memory) I'm seeing between 22GB and 26GB of ram used in total (max 1.625GB per thread), and I'm having to work some magic in linux to keep it running 15x Rosetta threads and 1 Folding at Home slot for my 1080Ti.I am kinda wanting just to double down on the current kit but i would hate the damper BOINC/Rosetta gets if suddenly i need more then 32gb of ram. I would have to limit the usage and i would hate having to do that. Given you all been running this a while i honestly am a bit clueless on what option to really go for. In regards to a 4x16gb kit i did find this.
Are you a human?
www.newegg.com
TLDR; I think you would end up running out of ram with 32GB, so 64 GB is a good choice. Worst case, we could help you find something to do with extra ram in Boinc (climateprediction.net uses 2-4GB per process if I remember correctly).
@deskjet, do you have enough RAM? (8 GB per running task, but you have certainly seen this in this thread already.)
In general, when I prepare a bunker for GPU tasks, I first suspend GPU computing, then [Allow new tasks], see to it that the respective quantity of tasks finished downloading, then suspend networking, then resume GPU computing. (This process can be automated and typically works out well unattended then.)
Amicable Numbers has tasks which may be quick or slow, but it also responds to work requests quickly and downloads quickly. Therefore these two extra steps of suspending GPU computing before and resuming after aren't necessary.
But more importantly: The server grants only 20 tasks in progress to a host, per GPU. (E.g., 60 tasks in progress for a triple GPU computer.) So the answer to your question of how many days to set the buffer size to: This doesn't matter all that much at Amicable Numbers, unless you had a very slow GPU. You won't get more than these 20 per GPU even with a large work buffer setting.
This is similar to a limit which SETI@home used to have. Their limit on tasks in progress was OK for computers with slow GPUs, but not for ones with modern GPUs. In addition, SETI@home had regular maintenance periods during which no new work could be fetched. Therefore, a few SETI@home users modified their boinc clients to show an inflated number of GPUs present when requesting work at the SETI server. This allowed them to have a reasonable amount of tasks in progress.
The required modification is arcane but not all that hard to do. The mentioned circle of SETI@home users were open about that they used this workaround, but kept the precise recipe to themselves, because inexperienced/ uncaring/ malicious users might make bad use of it.