RAM usage
26 Jan 2014 14:09:53 UTC
Topic 197358
(moderation:
Hi,
What is the influence of the amount of RAM on completing wu's? Going from a basic standpoint that you already have already RAM in reserve for basic OS operations and BOINC usage. Is for example 12 or 16 GB RAM going to make any difference in the duration of completing a file?
RAM usage
)
Yes it makes a difference, but no when you go above 8gb it is not helpful to Boinc. For the future I always put in 16gb when building a new machine, but right now it is overkill. The thing that make Boinc faster, up to a point, is faster ram, here is a wiki page showing a chart, scroll down to where it says JEDEC standard modules: http://en.wikipedia.org/wiki/DDR3_SDRAM
As you can see 2133 speed ram is almost 3 times as fast as 800 ram, that does make a difference in crunching speed. But costs also go up, often substantially, when going for higher speed ram. I buy what is on sale and 'settle', as I crunch for the long term not short term goals.
An intresting question. My
)
An intresting question.
My experience is: I run 12 Einstein wu's and 4-5 programs, but only a max of 80% of the 8GB ram is used. And I see no big difference in crunching speed when no other program runs.
The thing with the ram speed: yes, it is faster with faster ram, but not as much as one might expect just by looking to the frequencies. The reason is the cache memory, which should have the data ready in many many cases.
But maybe we have some hardware specialist here who can explain that with more background knowledge.
RE: An intresting
)
I agree, I also might have misled when I talked about one kind of memory being 3 times as fast as another, that is NOT directly related to how fast Boinc crunches!! It IS related as Alex said to how fast the memory is accessed though, some projects take better advantage of this then others.
Mikey and Alex got it pretty
)
Mikey and Alex got it pretty much right: the amount of RAM doesn't matter at all, as long as you have enough of it (no constant un/paging or "disk scratching"). The RAM frequency and timings do matter, at Einstein more so than for other projects. But we're talking about 5 - 10% here when comparing sane base configurations to "still affordable high end" stuff.
For a cruncher I'd currently use 2x4 GB (assuming you're not using socket 2011 with 4 RAM channels) of DDR3-1866 or 2133 with decent timings, as these are just a bit more expensive than baseline 1333 or 1600 modules. My main rig even uses DDR3-2400 with good timings.. but that did cost a bit more and is really only useful for POEM, where the GPU app was exceptionally demanding on the CPU and main memory bandwidth (if you ran a high end GPU). So this is not a general recommendation.
If you also crunch on the integrated GPU of current Intels or AMD APUs I wouldn't buy anything below DDDR3-2133.
MrS
Scanning for our furry friends since Jan 2002
RE: If you also crunch on
)
THX for that, MrS,
I took a look into the specs of the current APU's. For the i7 i found:
Quad-Core "Haswell" • Taktfrequenz: 3.50GHz, Turbo: 3.90GHz • TDP: 84W • Fertigung: 22nm • Interface: DMI, 5GT/s • L2-Cache: 4x 256kB • L3-Cache: 8MB shared • Stepping: C0 • Grafik: Intel HD Graphics 4600, 350-1200MHz, max. 3 Displays • PCIe-Lanes: 16x PCIe 3.0 • Sockel: 1150, max. 1 CPU • Memory Controller: Dual Channel PC3-12800U (DDR3-1600), 25.6GB/s, max. 32GB •
The newest A10 is rated DDR3-2133. So Intel is behind at the moment or the specs are wrong.
The expected 5-10% - can it be that the speed gain is higher for AMD (no Level 3 Cache) than for Intel (has L3 Cache)?
Alex
RE: RE: If you also
)
I don't know anymore, but a few years back this was a bad number "L2-Cache: 4x 256kB" as most cpu apps were bigger then that and resulted in lots of swapping, slowing down the crunching. 512kB was MUCH better, with the 256kB, or less, being an AMD Sempron or Intel Celeron chip.
@Alex: sorry, forgot to
)
@Alex: sorry, forgot to mention that. Itel is very conservative about their officially supported memory. Z-chipsets will let you set whatever you've got and whatever works. The world records with 3000+ MHz DDR3 are done with Intels :D
But cheaper chipsets will only let you set 1600 MHz as maximum, because they "don't support" overclocking. I heard it's not as straight on the AMD side, because getting 2400 MHz to work is pushing your luck or patience. Not really sure, though.
@Mikey: yeah, 256k L2 without L3 would be bad today. But no modern CPU has this configuration. Intel has their very fast 256k L2 caches coupled with a few MBs of reasonably fast L3, depending on the model. And AMD has their larger 1st and 2nd level caches (and only the biggest CPUs with an additional L3$).
MrS
Scanning for our furry friends since Jan 2002
RE: @Mikey: yeah, 256k L2
)
So the workunits auto roll over into L3 cache when they run out of L2 cache is what you are saying? If so then all is good again as both are MUCH faster then harddrive swapping!!
Just an idea: if ram costs is
)
Just an idea:
if ram costs is not a prime factor, one could use a virtual drive. AFAIK microsoft has not very clever routines to handle the existing ram, so why not use the ram as virtual drive for swapping? Should be at least 10 times faster than swapping to a SSD.
http://www.pcworld.com/article/260918/how_to_supercharge_your_pc_with_a_ram_disk.html
http://www.softperfect.com/products/ramdisk/
RE: So the workunits auto
)
Exactly - and that's what different cache levels are made for :)
There are still some details, e.g. AMD generally uses exclusive caches (the contents in each level are unique, data is moved between them and to get the total capacity you just add them all up) whereas Intel generally uses inclusive ones (L2 contains everything the L1 has as well, L3 includes L1 and L2). This gives Intel less net capacity but gives them performance, as it saves copies and moves between cache levels. But generally te above is true and automatically handled by the hardware.
Which is, BTW, a real issue for real time aplications on big CPUs with caches: the more complex the CPU and its caches are, the harder it gets for the programmer to predict when a task will be finished.
True, but before that there's the main memory, luckily! You can approximate it as another cache level.
I don't think so. Currently unused RAM is being used by Win as file cache (recently used ones) and since Vista for super fetch (often used ones).
If RAM contents are swapped to disk, despite RAM still being available, it's because you haven't used it since a long time. In that case.. why "cheat" around swapping to keep these pages in memory, when you could instead let Win use that memory for file cache & super fetch?
If RAM contents are swapped to disk because you're running out of RAM.. well, you need to swap pages to disk anyway. It wouldn't help to trigger this point earlier by reducing the amount of RAM available to Win via a RAM disk.
I can only see a point for RAM disks in very special cases. For example you've got a heavy I/O load on relatively small files, with lot's of writes (so caching isn't enough). In this case your application might be better off handling things in memory first and flushing to disk in a more efficient manner.. but that's not always an option. But here you have to specify the files to be kept in memory, which may not always be the best use of that RAM. Yes, that's why I primarily dislike RAM disks for general usage: they require a static content assignment, which I expect to be worse usage of that RAM space on average than automatic handling. Especially with SSDs, where some paging isn't that bad.
MrS
Scanning for our furry friends since Jan 2002