AFAIK microsoft has not very clever routines to handle the existing ram, so why not use the ram as virtual drive for swapping?
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).
MrS
Yeah Windows sure messed up the idea of 'free ram' didn't they? It took a whole new thought process when they started using ram as cache memory, and then automatically freeing it up when something else needs it. XP 'free ram' good, Vista and later 'free ram' bad! There IS still some 'free ram' so a program can start loading and then Windows gives it more as it needs it as it continues to load. You wouldn't want Windows to have to swap first as things could be really slow if your system were busy crunching or something.
As long as you have enough ram to avoid swapping more doesn't matter. If you're swapping more ram (or fewer WUs at a time) will give a large payout. I believe S6CasA has the ram usage of any E@H app; on my windows box they're using ~480MB each but have peaked at 660MB. For an 8 core box that's almost 5GB, so for a crunchbox 8 should be plenty but if you're trying to use it at the same time 16 is probably a better target.
Even on the crunchbox I'd be tempted to do 16 just to have headroom for future apps. A few years ago the apps only took ~250MB each. A new future version doubling it again is something I'd consider plausible over the lifetime of a box; and with DDR4 ram expected to start launching in volume this year DDR3 probably isn't going to fall in price a lot more.
and with DDR4 ram expected to start launching in volume this year DDR3 probably isn't going to fall in price a lot more.
That's true, but DDR3 prices have actually increased over the last year due to strong demand. If that demand shifts, prices might get back to where they used to be. Although historically superceeded RAM didn't drop much in price.
Anyway, I wouldn't want to buy RAM for future proofing crunching boxes now. It could be that the faster DDR4 modules drive the price markup on higher end DDR3 down, so in 2 years from now it wouldn't matter if you get DDR3-1600 or 2400. Or you might not need those 16 GB during the active crunching time of the machine at all. You could also run a project mix (recommended anyway) with some smaller tasks thrown in. This way you could still be fine with 8 GB even if Einstein required 1 GB per task. Which I suppose is not going to happen soon, since the analysis they're doing on this data is very time-consuming. They could easily expand the analysis rather than the data set to analyze if our CPUs became too fast.
Edit@Mikey: ehm.. no! Using RAM as file cache has been standard for operating system since a long time and is not usually shown to the user as "used memory". Since that would be pointless, always showing 99% in use. It's also not presented to programs as used, so it doesn't provide any hurdle to programs requesting memory. Super Fetch just uses the same unused RAM in the same way, but for a different purpose. Which I think is clearly superior to what we had, although I couldn't quantify the benefits.
I recently change RAM in two of my hosts:
1. PC1 -> AMD Phenom II X4 965 3.4GHz from DDR2-800 to DDR2-1066
2. PC2 -> AMD Phenom II 1090T 6-Core (3.4GHz) from DDR2-1066 to DDR3-1600 (new mb).
These changes has no impact in the performance of EaH CPU/GPU clients.
This is very strange, especially for the case of PC2 (DDR2-1066 to DDR3-1600) because I have noticed an overall performance boost of ~20-30% for tasks like compilation (which have a lot of memory usage).
Perhaps EaH clients do not depend a lot on memory speed.
Perhaps EaH clients do not depend a lot on memory speed.
That's unexpected. A possible reason would be significantly worse timings at the higher clock speeds. But then you probably wouldn't see a performance increase elsewhere, unless that was actually caused by the amount of memory.
On Intel CPUs there was a clear correlation between E@H performance and DRAM speed. The last proper measurements I've seen of this are a bit dated, but last spring when I upgraded my main memory from dual channel DDR3-1950 to DDR3-2400 I've seen about a ~10% performance increase for CPU tasks on my i7 3770K @ 4.1 GHz.
One could suspect the AMD cores would be so slow that DDR2-800 is enough to feed them.. but that's unlikely for your 6 core CPU. AMDs memory controller in those Phenom II chips didn't respond all that well to increased memory clocks, if I remember correctly. It would have been nice to see bandwidths & latency measurements before and after the change, to be sure. But since you changed the mainboard that would be inconvenient.
You could increase the memory controller clock speed. They usually reach ~2.8 GHz instead of the default 1.8 - 2.2 GHz. This increases the bandwidth and reduces the latency which can be extracted from the same memory and should hence increase E@H performance. Unless I'm totally wrong ;)
I recently change RAM in two of my hosts:
1. PC1 -> AMD Phenom II X4 965 3.4GHz from DDR2-800 to DDR2-1066
2. PC2 -> AMD Phenom II 1090T 6-Core (3.4GHz) from DDR2-1066 to DDR3-1600 (new mb).
These changes has no impact in the performance of EaH CPU/GPU clients.
This is very strange, especially for the case of PC2 (DDR2-1066 to DDR3-1600) because I have noticed an overall performance boost of ~20-30% for tasks like compilation (which have a lot of memory usage).
Perhaps EaH clients do not depend a lot on memory speed.
Just because you put faster memory in the machine doesn't always mean the machine is automatically going to use the faster speeds. Have you run a test to see the actual speed your ram is running at as opposed to what it can run at? I'm thinking your mb didn't 'auto adjust' itself to the new speeds it is seeing, but is instead just carrying on with the old settings in effect down clocking your ram.
Just because you put faster memory in the machine doesn't always mean the machine is automatically going to use the faster speeds.
Motherboards (or rather BIOSes) I have require you to change the XMP profile in the BIOS to use the full speed of the memory. This of course only applies if you bought memory with a XMP profile.
I try this tool (bandwidth) to measure the memory bandwidth and I notice that the PC with DDR3 has limited memory bandwidth compared to a DDR2 PC (also cpu cache bandwidth was limited).
After some searching I change some settings in the kernel (enable NUMA and Memory Compaction) and I get a performance boost ~50% on kernel compilation.
Also I get 400% more cpu cache bandwidth and 150% more memory bandwidth. The PC with DDR3 now has ~50% more memory bandwidth compared to DDR2 PC.
I will crunch some EaH WU's and I will measure the performance boost after these changes....
I try to change the same settings on a Intel-i7 PC with no effect.
It is very strange (to me) that the NUMA support will have such an impact in memory/cache bandwidth and overall performance of a single-cpu (multi-core) system.
I am not sure if this performance boost is a result of the memory compaction support only (also very strange).
I will try some kernel configurations with and without NUMA/MemCompaction and I will report the results...
RE: RE: AFAIK microsoft
)
Yeah Windows sure messed up the idea of 'free ram' didn't they? It took a whole new thought process when they started using ram as cache memory, and then automatically freeing it up when something else needs it. XP 'free ram' good, Vista and later 'free ram' bad! There IS still some 'free ram' so a program can start loading and then Windows gives it more as it needs it as it continues to load. You wouldn't want Windows to have to swap first as things could be really slow if your system were busy crunching or something.
As long as you have enough
)
As long as you have enough ram to avoid swapping more doesn't matter. If you're swapping more ram (or fewer WUs at a time) will give a large payout. I believe S6CasA has the ram usage of any E@H app; on my windows box they're using ~480MB each but have peaked at 660MB. For an 8 core box that's almost 5GB, so for a crunchbox 8 should be plenty but if you're trying to use it at the same time 16 is probably a better target.
Even on the crunchbox I'd be tempted to do 16 just to have headroom for future apps. A few years ago the apps only took ~250MB each. A new future version doubling it again is something I'd consider plausible over the lifetime of a box; and with DDR4 ram expected to start launching in volume this year DDR3 probably isn't going to fall in price a lot more.
RE: and with DDR4 ram
)
That's true, but DDR3 prices have actually increased over the last year due to strong demand. If that demand shifts, prices might get back to where they used to be. Although historically superceeded RAM didn't drop much in price.
Anyway, I wouldn't want to buy RAM for future proofing crunching boxes now. It could be that the faster DDR4 modules drive the price markup on higher end DDR3 down, so in 2 years from now it wouldn't matter if you get DDR3-1600 or 2400. Or you might not need those 16 GB during the active crunching time of the machine at all. You could also run a project mix (recommended anyway) with some smaller tasks thrown in. This way you could still be fine with 8 GB even if Einstein required 1 GB per task. Which I suppose is not going to happen soon, since the analysis they're doing on this data is very time-consuming. They could easily expand the analysis rather than the data set to analyze if our CPUs became too fast.
Edit@Mikey: ehm.. no! Using RAM as file cache has been standard for operating system since a long time and is not usually shown to the user as "used memory". Since that would be pointless, always showing 99% in use. It's also not presented to programs as used, so it doesn't provide any hurdle to programs requesting memory. Super Fetch just uses the same unused RAM in the same way, but for a different purpose. Which I think is clearly superior to what we had, although I couldn't quantify the benefits.
MrS
Scanning for our furry friends since Jan 2002
Hi, I recently change RAM
)
Hi,
I recently change RAM in two of my hosts:
1. PC1 -> AMD Phenom II X4 965 3.4GHz from DDR2-800 to DDR2-1066
2. PC2 -> AMD Phenom II 1090T 6-Core (3.4GHz) from DDR2-1066 to DDR3-1600 (new mb).
These changes has no impact in the performance of EaH CPU/GPU clients.
This is very strange, especially for the case of PC2 (DDR2-1066 to DDR3-1600) because I have noticed an overall performance boost of ~20-30% for tasks like compilation (which have a lot of memory usage).
Perhaps EaH clients do not depend a lot on memory speed.
RE: Perhaps EaH clients do
)
That's unexpected. A possible reason would be significantly worse timings at the higher clock speeds. But then you probably wouldn't see a performance increase elsewhere, unless that was actually caused by the amount of memory.
On Intel CPUs there was a clear correlation between E@H performance and DRAM speed. The last proper measurements I've seen of this are a bit dated, but last spring when I upgraded my main memory from dual channel DDR3-1950 to DDR3-2400 I've seen about a ~10% performance increase for CPU tasks on my i7 3770K @ 4.1 GHz.
One could suspect the AMD cores would be so slow that DDR2-800 is enough to feed them.. but that's unlikely for your 6 core CPU. AMDs memory controller in those Phenom II chips didn't respond all that well to increased memory clocks, if I remember correctly. It would have been nice to see bandwidths & latency measurements before and after the change, to be sure. But since you changed the mainboard that would be inconvenient.
You could increase the memory controller clock speed. They usually reach ~2.8 GHz instead of the default 1.8 - 2.2 GHz. This increases the bandwidth and reduces the latency which can be extracted from the same memory and should hence increase E@H performance. Unless I'm totally wrong ;)
MrS
Scanning for our furry friends since Jan 2002
RE: Hi, I recently change
)
Just because you put faster memory in the machine doesn't always mean the machine is automatically going to use the faster speeds. Have you run a test to see the actual speed your ram is running at as opposed to what it can run at? I'm thinking your mb didn't 'auto adjust' itself to the new speeds it is seeing, but is instead just carrying on with the old settings in effect down clocking your ram.
RE: Just because you put
)
Motherboards (or rather BIOSes) I have require you to change the XMP profile in the BIOS to use the full speed of the memory. This of course only applies if you bought memory with a XMP profile.
The new memories are running
)
The new memories are running at extreme profile (@1600MHz/9-9-9-27) and I noticed a performance boost of ~20-30% on several compilation tasks.
Hi, I try this tool
)
Hi,
I try this tool (bandwidth) to measure the memory bandwidth and I notice that the PC with DDR3 has limited memory bandwidth compared to a DDR2 PC (also cpu cache bandwidth was limited).
After some searching I change some settings in the kernel (enable NUMA and Memory Compaction) and I get a performance boost ~50% on kernel compilation.
Also I get 400% more cpu cache bandwidth and 150% more memory bandwidth. The PC with DDR3 now has ~50% more memory bandwidth compared to DDR2 PC.
I will crunch some EaH WU's and I will measure the performance boost after these changes....
I try to change the same settings on a Intel-i7 PC with no effect.
It is very strange (to me) that the NUMA support will have such an impact in memory/cache bandwidth and overall performance of a single-cpu (multi-core) system.
I am not sure if this performance boost is a result of the memory compaction support only (also very strange).
I will try some kernel configurations with and without NUMA/MemCompaction and I will report the results...