I've always run compcache on my ROMs when possible and I recently had the idea that setting the full amount of RAM to compcache could be an interesting test. The theory being that it may be slightly slower overall to have ramzswap compressing pages, but minfree would have heaps of memory to play with and apps would never quit, so multitasking would be faster as no apps are writing out to /data and then having to relaunch when I switch to them again.
I've set compcache with the init script like so
Code:
insmod /system/lib/compcache/ramzswap.ko;
rzscontrol /dev/block/ramzswap0 -i -d 98304;
busybox swapon /dev/block/ramzswap0;
And Linux reports the swap is working
Code:
# free
total used free shared buffers
Mem: 107332 105972 1360 0 8
Swap: 98296 48940 49356
Total: 205628 154912 50716
But no matter what I do, even if I launch a heap of apps, I can never get swap usage above the ~48Mb RAM seen here.
Furthermore, apps like Auto Killer and Free Memory only report I have 102Mb RAM total.
What's going on here? Why does Android think it only has 100Mb when free is reporting near on 200Mb total? Why does compcache only fill up halfway and no more? Does this mean that ~48Mb RAM is the "sweet spot" for compcache and any more is pointless?
The phone does feel faster when switching between apps but I could just be imagining it. Out of interest this is a 32B Magic running Dwang 1.17.1, which is basically just Donut AOSP with a faster framework and kernel.
Any ideas/help/suggestions would be appreciated?
Swap is never, ever, regarded by the system as real memory, hence why the system won't report it, since you really only have ~100 mb physical ram.
As to why you're only using 48, I believe that 48 is 1/2 96, so it probably means pages are being compressed 2:1, so the full memory is being compressed and dumped into RAM, and it only occupies 48 MB, leaving you 48 free for other processes.
Again, remember comcapche is swap, not real memory.
jubeh said:
Swap is never, ever, regarded by the system as real memory, hence why the system won't report it, since you really only have ~100 mb physical ram.
As to why you're only using 48, I believe that 48 is 1/2 96, so it probably means pages are being compressed 2:1, so the full memory is being compressed and dumped into RAM, and it only occupies 48 MB, leaving you 48 free for other processes.
Again, remember comcapche is swap, not real memory.
Click to expand...
Click to collapse
I don't think that that analysis is quite right....
First off, you MUST maintain SOME amount of real memory available... otherwise it'll crash in a spectacular way. I believe that the linux kernel itself may have a safety feature that maintains a certain minimum amount of physical ram available. There ARE certain things that the linux kernel will not be willing to swap, such as ITSELF.
Just imagine what would happen if the kernel swapped itself..... any attempt to do this wouldn't end well. Especially if it tried to swap its ENTIRE self since the kernel MUST be in memory in order for it to run.
There is also the swappiness setting... it controls the system's tendency to swap.
And finally, there is the possibility that you may simply not be starting enough processes to consume the full memory!
So imagine this; you have your compcache set for a certain size. It grows to that size and then finally, the kernel says "screw you, you can't have any more memory!" blows an error back to compcache, which complains back to the kernel "sorry, swap is screwed." Yep.... the kernel tells compcache which tells the kernel rather than the kernel just knowing.
You definitely don't want this happening.
Note: I can forsee some serious stability problems that this could result in related to the low memory process killer. Specifically, your compcache grows to its maximum allowed size, you start an application, the low memory process killer figures that you've got plenty of memory available, doesn't kill anything off, tries to start some application, crashes spectacularly when the kernel complains back that it doesn't have any memory. I don't know if this would happen with a stock low memory process killer, but definitely would with the swap hacks added....
lbcoder said:
Note: I can forsee some serious stability problems that this could result in related to the low memory process killer. Specifically, your compcache grows to its maximum allowed size, you start an application, the low memory process killer figures that you've got plenty of memory available, doesn't kill anything off, tries to start some application, crashes spectacularly when the kernel complains back that it doesn't have any memory. I don't know if this would happen with a stock low memory process killer, but definitely would with the swap hacks added....
Click to expand...
Click to collapse
92 MB of compcache doesn't really need 92MB of compcache... that's the point of being compcache.
Compcache file in RAM grows when cache gets stuffed inside compcache.
Setting a low swappiness will cause compcache to just swap what's needed.
And even with full compcache, in the end you end up having around 140 mb (or so) free ram. 92mb of compcache that takes like 50mb and 42 extra mb of normal ram.
I think this idea is great (I was just too lazy to try yet...). Instead of dalvik vm having to free up memory it can stuff some more mb in compcache. should be faster.
I didn't really think my post through... but I hope some of you understood some of the things I wanted to communicate xD
domenukk said:
92 MB of compcache doesn't really need 92MB of compcache... that's the point of being compcache.
Compcache file in RAM grows when cache gets stuffed inside compcache.
Setting a low swappiness will cause compcache to just swap what's needed.
And even with full compcache, in the end you end up having around 140 mb (or so) free ram. 92mb of compcache that takes like 50mb and 42 extra mb of normal ram.
I think this idea is great (I was just too lazy to try yet...). Instead of dalvik vm having to free up memory it can stuff some more mb in compcache. should be faster.
I didn't really think my post through... but I hope some of you understood some of the things I wanted to communicate xD
Click to expand...
Click to collapse
Interesting in theory, but if you actually read what I said, you would note that this is entirely IMPOSSIBLE and would crash spectacularly if not for and in some cases in SPITE of certain safety features built into the kernel.
Note: If you have 70 MB worth of data that CAN'T be swapped, that leaves 20 MB ***PEAK*** available to compcache.
It is neither fair nor sensible to think of all memory as being equal. Running processes ***MUST*** have REAL MEMORY.
A little off-topic, but this discussion (the possibility of REAL "compressed" memory) sparked a thought/question:
Would/could KSM* bring any benefit to Android? (Not sure if the KSM module can even compile/work on ARM)
I know KSM is normally used for detecting and sharing duplicate pages among KVM guests, but I wonder how many pages in a typical running Android installation are duplicated, and thus candidates for sharing/de-duplication.
*I can't posts links yet, so those that don't know what KSM is, will just have to google for it.
lbcoder said:
There ARE certain things that the linux kernel will not be willing to swap, such as ITSELF.
Click to expand...
Click to collapse
I was under the impression the kernel keeps itself in RAM and then reports free memory to the rest of the OS. This is why the phone physically has 192Mb RAM, but only reports 96Mb free (or 107Mb with RAM hack). Perhaps my understanding of Linux/Android memory reporting is not correct?
lbcoder said:
It is neither fair nor sensible to think of all memory as being equal. Running processes ***MUST*** have REAL MEMORY.
Click to expand...
Click to collapse
I think this is likely what is happening. Home, Phone, System and other processes with a low oom are refusing to swap out as they are still running. If the compcache allocation in RAM is dynamic as domenukk says, then those processes are occupying enough RAM that the ramzswap allocation can only grow to ~48Mb as I am seeing. I didn't think of this.
Nor have I tweaked swappiness. It's currently set to 60 (default) so I'd assume it's not too fussy with paging out. I will try playing with this at 10 and 100 to see if I can force anything more into swap or if it's less willing to swap.
brainbone said:
*I can't posts links yet, so those that don't know what KSM is, will just have to google for it.
Click to expand...
Click to collapse
I do not know either, but here are some links people may wish to look at
http://fedoraproject.org/wiki/Features/KSM
http://lwn.net/Articles/306704/
http://lwn.net/Articles/330589/
http://www.linux-kvm.com/content/using-ksm-kernel-samepage-merging-kvm
Ok I am not at all experienced in this area but this is just a suggestion. You say that you can only get 48mb of swap to be compressed at a time. If those 48mb were uncompressed, then that would occupy the 98mb you alloted to it. If you set the cc at say 128mb, then in (my) theory up to 64mb of it the actual ram would be used. I don't know how much sense I'm making but not sure exactly how to explain it. If you don't get it I'll try explaining my logic a little more in depth.
mejorguille said:
If you set the cc at say 128mb, then in (my) theory up to 64mb of it the actual ram would be used.
Click to expand...
Click to collapse
It appears you are right. Even with swappiness set to 100 and opening every app on my phone I'm not able to fill more than half of 128Mb compcache before minfree stats gracefully closing processes:
Code:
/opt/android-sdk-linux_86/tools$ ./adb shell free
total used free shared buffers
Mem: 107332 105956 1376 0 32
Swap: 131064 65520 65544
Total: 238396 171476 66920
Super Jamie said:
It appears you are right. Even with swappiness set to 100 and opening every app on my phone I'm not able to fill more than half of 128Mb compcache before minfree stats gracefully closing processes:
Code:
/opt/android-sdk-linux_86/tools$ ./adb shell free
total used free shared buffers
Mem: 107332 105956 1376 0 32
Swap: 131064 65520 65544
Total: 238396 171476 66920
Click to expand...
Click to collapse
I rock=p
So what's your performance like, compared to say 32mb cc or no cc at all?
It's different but I like it.
There is sometimes a slight (<2 second) pause when launching a new app (I assume this is compcache compressing old pages to swap to make way for the new app) however once the app is up and running, it almost never "exits" so switching between previously-launched apps is noticeably faster than without compcache. I run HelixLauncher Donut and it's never closed and re-launched while I've been trying this, however it did sometimes with 32Mb cc and quite often without cc at all.
I had 5 day uptime last week with 96Mb compcache (rebooted whilst testing another app for a friend) so I don't think stability is an issue. The CM wiki indicates performance with cc is better upon boot then gradually declines, even if that is the case, rebooting my phone once a week is no big issue.
Unless I run into any major issues, I'll be keeping my phone with large compcache
Super Jamie said:
I was under the impression the kernel keeps itself in RAM and then reports free memory to the rest of the OS. This is why the phone physically has 192Mb RAM, but only reports 96Mb free (or 107Mb with RAM hack). Perhaps my understanding of Linux/Android memory reporting is not correct?
I think this is likely what is happening. Home, Phone, System and other processes with a low oom are refusing to swap out as they are still running. If the compcache allocation in RAM is dynamic as domenukk says, then those processes are occupying enough RAM that the ramzswap allocation can only grow to ~48Mb as I am seeing. I didn't think of this.
Nor have I tweaked swappiness. It's currently set to 60 (default) so I'd assume it's not too fussy with paging out. I will try playing with this at 10 and 100 to see if I can force anything more into swap or if it's less willing to swap.
I do not know either, but here are some links people may wish to look at
http://fedoraproject.org/wiki/Features/KSM
http://lwn.net/Articles/306704/
http://lwn.net/Articles/330589/
http://www.linux-kvm.com/content/using-ksm-kernel-samepage-merging-kvm
Click to expand...
Click to collapse
ksm sound cool. As every app runs inside dalvik vm. Not sure though... somebody should ask cyanogen
I am happy thatlarge compcache works out so well for you.
BTW doesn't cyanogenmod 5 count the whole swap as real memory since test4 or so?
Oh and overclocking ondemand to as high as possible will speed up app opening and switching a lot while on compcache
domenukk said:
doesn't cyanogenmod 5 count the whole swap as real memory since test4 or so?
Click to expand...
Click to collapse
It's counted as available "swap" memory, but not "real" ram. Pages stored in "real" ram (memory that the cpu can directly execute code in) still need to be freed up (moved to swap) before previously swapped pages can be moved back in to "real" memory to be executed.
domenukk said:
ksm sound cool.
Click to expand...
Click to collapse
The beauty of KSM is that it does no "swapping". It simply combines 4KB pages that are identical -- so instead of two identical 4KB pages using 8KB of ram, they only take 4KB. The code is executed in place.
KSM would, however, still require swap. If at any time a virtual shared 4KB page is written to, it needs to be copied to a free page to avoid corrupting the virtual page it was a duplicate of before the write. Swap is needed to ensure that there will always be enough available free pages when this happens.
domenukk said:
As every app runs inside dalvik vm.
Click to expand...
Click to collapse
KSM is not dependent on a VM, but the existence of VMs (java or otherwise) increases the likelihood of duplicate pages.
domenukk said:
Not sure though... somebody should ask cyanogen
Click to expand...
Click to collapse
I'd certainly be interested in cyanogen's thoughts on this, but I'm sure there are others that would be able to chime in as well.
Relevant excerpt from kernel.org/doc/ols/2009/ols2009-pages-19-28.pdf
KSM and embedded
KSM is suitable to be run on embedded systems too; the important thing is not to register in KSM regions that won’t likely have equal pages. For each virtual page scanned, KSM has to allocate some rmap_item and tree_item, so while these allocations are fairly small, they can be noticeable if lots of virtual areas are scanned for no good.
Furthermore, these KSM internal rmap/tree data structures are not allocated in high memory. To avoid early out of memory conditions, it is especially important to limit the amount of lowmem allocated on highmem 32bit systems that might have more than 4GB of memory, but these shouldn’t fit in the embedded category in the first place.
Click to expand...
Click to collapse
Super Jamie said:
I was under the impression the kernel keeps itself in RAM and then reports free memory to the rest of the OS. This is why the phone physically has 192Mb RAM, but only reports 96Mb free (or 107Mb with RAM hack). Perhaps my understanding of Linux/Android memory reporting is not correct?
Click to expand...
Click to collapse
Memory reporting is a tricky thing.
But in general, when RAM is allocated to HARDWARE, it is NOT REPORTED.
The chunk of the 192 that is not reported is assigned PRIMARILY to the RADIO. Another chunk is assigned to the GPU. Still more is assigned to the AUDIO HARDWARE.
The part used by the kernel itself IS reported. The most trivial empirical evidence to prove this to you is that when you change KERNELS, it doesn't affect the total system memory, despite your KNOWING that different kernels use different amounts of RAM.
Another bit of empirical proof is that activating COMPCACHE does not reduce the total physical ram reported. And yes, COMPCACHE is part of the kernel...
Now here's another concept of crash and burn: IF everything in memory could be cached in compcache, then what would stop compcache from caching compcache in a horrible infinite memory sucking loop? That would be very very bad, LOL.
I think this is likely what is happening. Home, Phone, System and other processes with a low oom are refusing to swap out as they are still running. If the compcache allocation in RAM is dynamic as domenukk says, then those processes are occupying enough RAM that the ramzswap allocation can only grow to ~48Mb as I am seeing. I didn't think of this.
Click to expand...
Click to collapse
NOW you're getting the idea! Running processes, kernel, etc., all need physical RAM (though actually those processes you mention most definitely CAN be swapped...), and therefore you CAN'T make the ENTIRE RAM into compcache!
Nor have I tweaked swappiness. It's currently set to 60 (default) so I'd assume it's not too fussy with paging out. I will try playing with this at 10 and 100 to see if I can force anything more into swap or if it's less willing to swap.
Click to expand...
Click to collapse
Just beware of possible crash-and-burn
Super Jamie said:
It appears you are right. Even with swappiness set to 100 and opening every app on my phone I'm not able to fill more than half of 128Mb compcache before minfree stats gracefully closing processes:
Code:
/opt/android-sdk-linux_86/tools$ ./adb shell free
total used free shared buffers
Mem: 107332 105956 1376 0 32
Swap: 131064 65520 65544
Total: 238396 171476 66920
Click to expand...
Click to collapse
That doesn't actually follow from what you've posted here.
What follows is that 105956-(65520/2)=73196 of what occupies your memory can't be swapped (kernel, running processes, etc.).
lbcoder said:
What follows is that 105956-(65520/2)=73196 of what occupies your memory can't be swapped (kernel, running processes, etc.).
Click to expand...
Click to collapse
Do you know why swap constantly "settles" at almost exactly half usage regardless of what size compcache I set?
For example, I set 64Mb compcache yesterday and rebooted. Just using my phone normally (browser, genie widget, music) I have this:
Code:
total used free shared buffers
Swap: 63992 32096 31896
Does this mean I am effectively reducing the amount of RAM the phone has for the kernel, "foreground app", "visible app" and "secondary servers" (to use the minfree terms), whilst at the same time allowing more "hidden app" and lower processes to swap out instead of terminating gracefully?
This disturbs me
-------------------------------------
Sent via the XDA Tapatalk App
brainbone said:
I'd certainly be interested in cyanogen's thoughts on this, but I'm sure there are others that would be able to chime in as well.
Click to expand...
Click to collapse
He doesn't know much about it, yet. But he seems inerested.
Attached a short conversation over #twitter
# Dominik domenuk
@cyanogen Is ksm any good 2 save RAM? probably not - you would have already done it http://lwn.net/Articles/329123/
# Steve Kondik cyanogen
@domenuk I don't know too much about it, I think its meant for sharing between distinct virtual machines
@domenuk it could have a lot of potential though
# Dominik domenuk
@cyanogen basically yes. But he states its also for normal apps. I have no idea to what extend android apps have similar memory, though...
# Steve Kondik cyanogen
@domenuk a lot, Android is all about IPC and shared memory. I wouldn't be surprised if the Dalvik people are thinking about it.
Click to expand...
Click to collapse
Here is a way to make Android use more than 50 percent of a swap partition. I am not sure how it will act with compcache... Should be the same... Run the following commands from a terminal or add them following to your userinit.sh file:
Code:
su
echo 80 > /proc/sys/vm/swappiness
echo 150 > /proc/sys/vm/vfs_cache_pressure
!!WARNING!! - Messing with VM settings can cause data loss and system instability... Not liable for damages...
Using the above the "free" output is:
Code:
total used free shared buffers
Mem: 97708 95168 2540 0 356
Swap: 125296 88756 36540
Total: 223004 183924 39080
While we are at it... If anyone is willing... they can try this as well:
echo 1 > /proc/sys/vm/oom_kill_allocating_task
Reference : Article to Linux Insight
Been having pretty good results with it...
Here is a link to the rest of the vm settings...
Linux Insight article listing vm settings
L8r
hi guys, just wondering if Jit is somehow related to linpack.
Because I noticed that every custom for my phone scores worse results with "jit enabled"
but jit shouldn't enabled by default on froyo?
JIT allows code to be executed outside of the dalvikVM, by doing so code is processed directly on the hardware and is faster because there is no overhead of virtualization. JIT has proved to increase performance however in requires an allocation of memory and most configurations set it at 24-32MB .
Since you give no details of your phone or if you are overclocking or using a particular govenor, your question has no definitive answer except that you are using a configuration that produces poor performance. JIT has proven to give phones better performance but when mixing in swap partions and memory management can cause the system to crawl.
nitro-ale said:
hi guys, just wondering if Jit is somehow related to linpack.
Because I noticed that every custom for my phone scores worse results with "jit enabled"
but jit shouldn't enabled by default on froyo?
Click to expand...
Click to collapse
Phonekenstein said:
JIT allows code to be executed outside of the dalvikVM, by doing so code is processed directly on the hardware and is faster because there is no overhead of virtualization. JIT has proved to increase performance however in requires an allocation of memory and most configurations set it at 24-32MB .
Since you give no details of your phone or if you are overclocking or using a particular govenor, your question has no definitive answer except that you are using a configuration that produces poor performance. JIT has proven to give phones better performance but when mixing in swap partions and memory management can cause the system to crawl.
Click to expand...
Click to collapse
thanks for your answer, I have a huawei u8150, no overclock, no ext or swap partition on my sd.
Stock rom scores 3.7 mflops 1.7 secs and cust rom with jit scores 2.4 mflops in 2.7 secs... seems strange
I've been seeing This CM7 and wanted to know what this was?!
I'm always reading how people say that xxMB is the best size.
But what is it? And how it works?
What's it for? What's the right size for a Galaxy S?
VeryCoolAlan said:
I've been seeing This CM7 and wanted to know what this was?!
I'm always reading how people say that xxMB is the best size.
But what is it? And how it works?
What's it for? What's the right size for a Galaxy S?
Click to expand...
Click to collapse
This setting is simply the maximum amount of heap space (read: memory) a single instance of the Dalvik VM (read: application) can obtain.
Don't read anything into the "benchmarks" performed. This setting should have little effect on overall system performance. The only scenario where it would be beneficial to increase the maximum heap size would be if you have an application that is very close to using up all of its available heap space, which would force it to run garbage collection frequently, which would use up CPU cycles. It is possible that lowering the maximum heap size could be beneficial in that it might prevent an application from obtaining more memory than it needs (by forcing it to garbage collect sooner), but that all depends on how the Dalvik VM is implemented and is really beyond my knowledge.
The heap (in Java) stores dynamically allocated variables, such as objects. Like Rueben_ said, when the heap is running low on memory, the JVM will run garbage collection. Garbage collection uses processing cycles, which will slow down your phone. In case you're wondering, the other place in memory where things are stored is called the stack, which stores arguments and parameters.
Oh okay
So how do you how much memory to put it for at Max?
Sent from my SGH-T959 using xda premium
mrinehart93 said:
In case you're wondering, the other place in memory where things are stored is called the stack, which stores arguments and parameters.
Click to expand...
Click to collapse
I do believe the stack is reserved for storage of CPU data (typically, the contents of its registers) before it switches toanother context (or thread, which means loading a different set of data in its registers).
The memory zone where data from user apps (and the Dalvik VM is one of those, albeit a über-app of sorts since it runs all the other apps) is stored is rather just called "the heap". At least that's the way I learned it
daxdax89 said:
So is it better to put higher or lower heap size for games?
Click to expand...
Click to collapse
There is no right answer for this as you should have gathered from reading the answers you were given.
It all depends on how you use your phone.
Try out different settings with games and you'll see which suits your usage best.
Go on - Live on the edge a little lol...
daxdax89 said:
So is it better to put higher or lower heap size for games?
Click to expand...
Click to collapse
For the Vibrant if you use big apps then put the VM at like 128 or or something like that.
If you have small flashlight apps then use it at 32-96
Sent from my SAMSUNG-SGH-I747 using xda premium
Say, this seems to be a "magic bullet" for devices with modest memory sizes to keep apps from swamping the available memory. For my 1 GB Lollipop device, setting dalvik.vm.heapsize to "96m" improved the performance of the device. I was able to cycle between many memory-heavy applications without getting bogged down.
Be careful! Forgetting the "m" at the end, or setting the number too low will send your device into an endless garbage-collection, essentially locking you out. I set it too low myself, had to pop the battery, then quickly raise the value after reboot (before the "system_server" app got itself above the limit I had set), rebooting again to apply the new value. For a fair indication of a lower limit, look at the memory consumption of the "system_server" app. Set it above that to give it some headroom. Make sure you have a recovery plan before messing with this value!
VM heap size on my ROM is set (by default I assume, since I haven't mess with it) to 128. This seems a bit much to me, since I only use maybe 2 or 3 large apps. Is this fine or would I get more benefit from dropping it down?
48 works just fine but 64 is preferred.
Sent from my LG-P999
OK I'll give it a shot. I just figured 128 was a waste of resources
the vm is how much memory is reserved for each app that is running.
if you have 250ish memory available for processes to use, with 64mb VM, you will be able to keep 4 programs in memory (roughly)
with 128mb VM, that is cut down to about 2.
Honestly, unless you have a program that is crashing with lower VM (usually an intense game or trying to record HD video) 48mb is most likely your best bet. it will allow 5 ish apps to stay in memory, which will cut down load times and give you the best experience.
ive been out of the android bubble for a bit, but 128mb vm sounds insanely unnecessary.