frontswap/zcache: xvmalloc discussion
Subsystems (2)
Dan, Nitin,
I have been experimenting with the frontswap v4 patches and the latest zcache in the mainline drivers/staging. There is a particular issue I'm seeing when using pages of different compressibilities.
When the pages compress to less than PAGE_SIZE/2, I get good compression and little external fragmentation in the xvmalloc pool. However, when the pages have a compressed size greater than PAGE_SIZE/2, it is a very different story. Basically, because xvmalloc allocations can't span multiple pool pages, grow_pool() is called on each allocation, reducing the effective compression (total_pages_in_frontswap / total_pages_in_xvmalloc_pool) to 0 and drastically increasing external fragmentation to up to 50%.
The likelihood that the size of a compressed page is greater than PAGE_SIZE/2 is high, considering that lzo1x-1 sacrifices compressibility for speed. In my experiments, pages of English text only compressed to 75% of their original size with 1zo1x-1.
In order to calculate the effective compression of frontswap, you need the number of pages stored by frontswap, provided by frontswap's curr_pages sysfs attribute, and the number of pages in the xvmalloc pool. There isn't a sysfs attribute for this, so I made a patch that creates a new zv_pool_pages_count attribute for zcache that provides this value (patch is in a follow-up message). I have also included my simple test program at the end of this email. It just allocates and stores random pages of from a text file (in my case, a text file of Moby Dick).
The real problem here is compressing pages of size x and storing them in a pool that has "chunks", if you will, also of size x, where allocations can't span multiple chunks. Ideally, I'd like to address this issue by expanding the size of the xvmalloc pool chunks from one page to four pages (I can explain why four is a good number, just didn't want to make this note too long).
After a little playing around, I've found this isn't entirely trivial to do because of the memory mapping implications; more specifically the use of kmap/kunamp in the xvmalloc and zcache layers. I've looked into using vmap to map multiple pages into a linear address space, but it seems like there is a lot of memory overhead in doing that.
Do you have any feedback on this issue or suggestion solution?
Signature · 84 lines
Seth Jennings
Linux on Power Virtualization
IBM Linux Technology Center
=================
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#define SIZE_OF_PAGE_IN_BYTES 4096
#define NUM_OF_PAGES_PER_MB (1024*1024/SIZE_OF_PAGE_IN_BYTES)
void usage();
int
main(int argc, char * argv[])
{
int mbs, numpages, i;
char *mypage;
int rc, len, pos;
FILE *textfile;
if(argc < 3)
{
printf("usage: %s numMBs file\n",argv[0]);
return 1;
}
mbs = atoi(argv[1]);
numpages = NUM_OF_PAGES_PER_MB * mbs;
printf("Allocating %d MB (%d pages) of data...\n",mbs,numpages);
textfile = fopen(argv[2],"r");
if(textfile == NULL)
{
perror("failed to open text file");
exit(1);
}
/* get file length */
fseek(textfile,0,SEEK_END);
len = ftell(textfile);
for (i=0; i < numpages; i++) {
if(!(i%100))
{
if(i)
printf("\033[F\033[J");
printf("%d numpages allocated\n", i);
}
mypage = malloc(SIZE_OF_PAGE_IN_BYTES);
if(!mypage)
{
perror("malloc()");
return 1;
}
/* start at (pusedo) random location in file */
pos = rand() % (len - SIZE_OF_PAGE_IN_BYTES);
fseek(textfile,pos,SEEK_SET);
rc = fread(mypage,SIZE_OF_PAGE_IN_BYTES,1,textfile);
if(!rc)
{
perror("read()");
return 1;
}
}
printf("complete\n\npress any key to end program");
getchar();
return 0;
}
=================
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>PATCH 1/1 Add zv_pool_pages_count to zcache sysfs
This attribute returns the number of pages currently in use by the xvmalloc pool used to store persistent pages compressed by zcache. This attribute, in combination with the curr_pages attribute of frontswap, can be used to calculate the effective compression of frontswap (i.e. zv_pool_pages_count/curr_pages)
--- drivers/staging/zcache/zcache.c | 20 ++++++++++++++++++++ 1 files changed, 20 insertions(+), 0 deletions(-)
drivers/staging/zcache/zcache.c +20 -0
@@ -684,6 +684,23 @@ static struct {struct xv_pool *xvpool; } zcache_client; +#ifdef CONFIG_SYSFS +static int zv_show_pool_pages_count (char *buf) +{ +char *p = buf; +unsigned long numpages; + +if (zcache_client.xvpool ==NULL ) + p +=sprintf (p,"%d\n" ,0 ); +else { + numpages =xv_get_total_size_bytes (zcache_client.xvpool); + p +=sprintf (p,"%lu\n" , numpages >>PAGE_SHIFT ); + } + +return p - buf; +} +#endif +/* * Tmem operations assume the poolid implies the invoking client. * Zcache only has one client (the kernel itself), so translate
@@ -1130,6 +1147,8 @@ ZCACHE_SYSFS_RO_CUSTOM(zbud_unbuddied_list_counts, zbud_show_unbuddied_list_counts);ZCACHE_SYSFS_RO_CUSTOM (zbud_cumul_chunk_counts, zbud_show_cumul_chunk_counts); +ZCACHE_SYSFS_RO_CUSTOM (zv_pool_pages_count, + zv_show_pool_pages_count);static struct attribute *zcache_attrs[] = { &zcache_curr_obj_count_attr.attr,
@@ -1160,6 +1179,7 @@ static struct attribute *zcache_attrs[] = { &zcache_aborted_shrink_attr.attr, &zcache_zbud_unbuddied_list_counts_attr.attr, &zcache_zbud_cumul_chunk_counts_attr.attr, + &zcache_zv_pool_pages_count_attr.attr, NULL, };
Signature · 8 lines
1.7.0.4 -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Re: [PATCH] Add zv_pool_pages_count to zcache sysfs
Quoted 1 line · Seth Jennings
+#ifdef
CONFIG_SYSFS
There are a couple of #ifdef CONFIG_SYSFS blocks in zcache.c already. Could this go inside one of those instead of being off by itself?
drivers/staging/zcache/zcache.c static struct { · line 695
Quoted 8 lines
drivers/staging/zcache/zcache.c +8
-0+static int zv_show_pool_pages_count (char *buf) +{ +char *p = buf; +unsigned long numpages; + +if (zcache_client.xvpool ==NULL ) + p +=sprintf (p,"%d\n" ,0 ); +else {
^^ That's probably a good spot to include brackets. They don't take up any more lines, and it keeps folks from introducing bugs doing things like:
if (zcache_client.xvpool == NULL)
p += sprintf(p, "%d\n", 0);
bar();
else {drivers/staging/zcache/zcache.c static struct { · line 698
Quoted 3 lines
drivers/staging/zcache/zcache.c +3
-0+ numpages =xv_get_total_size_bytes (zcache_client.xvpool); + p +=sprintf (p,"%lu\n" , numpages >>PAGE_SHIFT ); + }
In this case 'numpages' doesn't actually store a number of pages; it stores a number of bytes. I'd probably rename it.
Also 'numpages' is an 'unsigned long' while xv_get_total_size_bytes() returns a u64. 'unsigned long' is only 32-bits on 32-bit architectures, so it's possible that large buffers sizes could overflow. The easiest way to fix this is probably to just make 'numpages' a u64.
-- Dave
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 5 lines
From: Seth Jennings [mailto:[email protected]]
Cc: Dan Magenheimer; Nitin Gupta; Robert Jennings; Brian King; Greg Kroah-Hartman
Subject: frontswap/zcache: xvmalloc discussionDan, Nitin,
Hi Seth --
Thanks for your interest in frontswap and zcache!
Quoted 13 lines
I have been experimenting with the frontswap v4 patches and the latest zcache in the mainline
drivers/staging. There is a particular issue I'm seeing when using pages of different compressibilities.When the pages compress to less than PAGE_SIZE/2, I get good compression and little external fragmentation in the xvmalloc pool. However, when the pages have a compressed size greater than PAGE_SIZE/2, it is a very different story. Basically, because xvmalloc allocations can't span multiple pool pages,
grow_pool()is called on each allocation, reducing the effective compression (total_pages_in_frontswap/total_pages_in_xvmalloc_pool) to 0 and drastically increasing external fragmentation to up to 50%.
Quoted 4 lines
The likelihood that the size of a compressed page is greater than PAGE_SIZE/2 is high, considering that lzo1x-1 sacrifices compressibility for speed. In my experiments, pages of English text only compressed to 75% of their original size with 1zo1x-1.
Wow, I'm surprised to hear that. I suppose it is very workload dependent, but I agree that consistently poor compression can create issues for frontswap.
Quoted 10 lines
In order to calculate the effective compression of frontswap, you need the number of pages stored by frontswap, provided by frontswap's
curr_pagessysfs attribute, and the number of pages in the xvmalloc pool. There isn't a sysfs attribute for this, so I made a patch that creates a newzv_pool_pages_countattribute for zcache that provides this value (patch is in a follow-up message). I have also included my simple test program at the end of this email. It just allocates and stores random pages of from a text file (in my case, a text file of Moby Dick).
Quoted 6 lines
The real problem here is compressing pages of size x and storing them in a pool that has "chunks", if you will, also of size x, where allocations can't span multiple chunks. Ideally, I'd like to address this issue by expanding the size of the xvmalloc pool chunks from one page to four pages (I can explain why four is a good number, just didn't want to make this note too long).
Nitin is the expert on compression and xvmalloc... I mostly built on top of his earlier work... so I will wait for him to comment on compression and xvmalloc issues.
BUT... I'd be concerned with increasing the pool chunk, at least without a fallback. When memory is constrained, finding chunks in the kernel of even two consecutive pages might be a challenge, let alone four. Since frontswap only is invoked if swapping is occurring, memory is definitely already constrained.
If it is possible to modify xvmalloc (or possibly the pool creation calls from zcache) to juggle multiple pools, one with chunkorder==2, one with chunkorder==1, and one with chunkorder=0, with a fallback sequence if a higher chunkorder is not available, might that be helpful? Still I worry that the same problems might occur because the higher chunkorders might never be available after some time passes.
Quoted 7 lines
After a little playing around, I've found this isn't entirely trivial to do because of the memory mapping implications; more specifically the use of kmap/kunamp in the xvmalloc and zcache layers. I've looked into using vmap to map multiple pages into a linear address space, but it seems like there is a lot of memory overhead in doing that.
Do you have any feedback on this issue or suggestion solution?
One neat feature of frontswap (and the underlying Transcendent Memory definition) is that ANY PUT may be rejected**. So zcache could keep track of the distribution of "zsize" and if the number of pages with zsize>PAGE_SIZE/2 greatly exceeds the number of pages with "complementary zsize", the frontswap code in zcache can reject the larger pages until balance/sanity is restored.
Might that help? If so, maybe your new sysfs value could be replaced with the ratio (zv_pool_pages_count/frontswap_curr_pages) and this could be _writeable_ to allow the above policy target to
be modified at runtime. Even better, the fraction could be
represented by number-of-bytes ("target_zsize"), which could default to something like (3*PAGE_SIZE)/4... if the ratio above exceeds target_zsize and the zsize of the page-being-put exceeds target_zsize, then the put is rejected.
Thanks,
Dan
** The "put" shouldn't actually be rejected outright... it should be converted to a "flush" so that, if a previous put was performed for the matching handle, the space can be reclaimed. (Let me know if you need more explanation of this.)
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 9 lines · Dan Magenheimer
Quoted 5 lines
From: Seth Jennings [mailto:[email protected]]
Cc: Dan Magenheimer; Nitin Gupta; Robert Jennings; Brian King; Greg Kroah-Hartman
Subject: frontswap/zcache: xvmalloc discussionDan, Nitin,
Hi Seth --
Thanks for your interest in frontswap and zcache!
Thanks for your quick response!
Quoted 14 lines
Quoted 14 lines
I have been experimenting with the frontswap v4 patches and the latest zcache in the mainline
drivers/staging. There is a particular issue I'm seeing when using pages of different compressibilities.When the pages compress to less than PAGE_SIZE/2, I get good compression and little external fragmentation in the xvmalloc pool. However, when the pages have a compressed size greater than PAGE_SIZE/2, it is a very different story. Basically, because xvmalloc allocations can't span multiple pool pages,
grow_pool()is called on each allocation, reducing the effective compression (total_pages_in_frontswap/total_pages_in_xvmalloc_pool) to 0 and drastically increasing external fragmentation to up to 50%.
Quoted 9 lines
Quoted 4 lines
The likelihood that the size of a compressed page is greater than PAGE_SIZE/2 is high, considering that lzo1x-1 sacrifices compressibility for speed. In my experiments, pages of English text only compressed to 75% of their original size with 1zo1x-1.
Wow, I'm surprised to hear that. I suppose it is very workload dependent, but I agree that consistently poor compression can create issues for frontswap.
Yes, I was surprised as well with how little it compressed. I guess I'm used to gzip level compression, which was around 50% on the same data set.
Quoted 17 lines
Quoted 10 lines
In order to calculate the effective compression of frontswap, you need the number of pages stored by frontswap, provided by frontswap's
curr_pagessysfs attribute, and the number of pages in the xvmalloc pool. There isn't a sysfs attribute for this, so I made a patch that creates a newzv_pool_pages_countattribute for zcache that provides this value (patch is in a follow-up message). I have also included my simple test program at the end of this email. It just allocates and stores random pages of from a text file (in my case, a text file of Moby Dick).Quoted 6 lines
The real problem here is compressing pages of size x and storing them in a pool that has "chunks", if you will, also of size x, where allocations can't span multiple chunks. Ideally, I'd like to address this issue by expanding the size of the xvmalloc pool chunks from one page to four pages (I can explain why four is a good number, just didn't want to make this note too long).
Quoted 4 lines
Nitin is the expert on compression and xvmalloc... I mostly built on top of his earlier work... so I will wait for him to comment on compression and xvmalloc issues.
Yes, I do need Nitin to weigh in on this since any changes to the xvmalloc code would impact zcache and zram.
Quoted 6 lines
BUT... I'd be concerned with increasing the pool chunk, at least without a fallback. When memory is constrained, finding chunks in the kernel of even two consecutive pages might be a challenge, let alone four. Since frontswap only is invoked if swapping is occurring, memory is definitely already constrained.
Quoted 8 lines
If it is possible to modify xvmalloc (or possibly the pool creation calls from zcache) to juggle multiple pools, one with chunkorder==2, one with chunkorder==1, and one with chunkorder=0, with a fallback sequence if a higher chunkorder is not available, might that be helpful? Still I worry that the same problems might occur because the higher chunkorders might never be available after some time passes.
To avoid the problem with getting one large set (up to 4 pages) of contiguous space, I'm looking into using vm_map_ram() to map chunks that are multiple noncontiguous pages into a single contiguous address space. I don't know what the overhead is yet.
I do like the idea of having a few pools with different chunk sizes.
Quoted 8 lines
Quoted 7 lines
After a little playing around, I've found this isn't entirely trivial to do because of the memory mapping implications; more specifically the use of kmap/kunamp in the xvmalloc and zcache layers. I've looked into using vmap to map multiple pages into a linear address space, but it seems like there is a lot of memory overhead in doing that.
Do you have any feedback on this issue or suggestion solution?
Quoted 8 lines
One neat feature of frontswap (and the underlying Transcendent Memory definition) is that ANY PUT may be rejected**. So zcache could keep track of the distribution of "zsize" and if the number of pages with zsize>PAGE_SIZE/2 greatly exceeds the number of pages with "complementary zsize", the frontswap code in zcache can reject the larger pages until balance/sanity is restored.
Might that help?
We could do that, but I imagine that would let a lot of pages through on most workloads. Ideally, I'd like to find a solution that would capture and (efficiently) store pages that compressed to up to 80% of their original size.
Quoted 9 lines
If so, maybe your new sysfs value could be
replaced with the ratio (zv_pool_pages_count/frontswap_curr_pages) and this could be _writeable_ to allow the above policy target to
be modified at runtime. Even better, the fraction could be
represented by number-of-bytes ("target_zsize"), which could default to something like (3*PAGE_SIZE)/4... if the ratio above exceedstarget_zsizeand the zsize of the page-being-put exceedstarget_zsize, then the put is rejected.
Quoted 7 lines
Thanks,
Dan** The "put" shouldn't actually be rejected outright... it should be converted to a "flush" so that, if a previous put was performed for the matching handle, the space can be reclaimed. (Let me know if you need more explanation of this.)
Thanks again for your reply, Dan. I'll explore this more next week.
--
Seth
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 7 lines
Quoted 7 lines
One neat feature of frontswap (and the underlying Transcendent Memory definition) is that ANY PUT may be rejected**. So zcache could keep track of the distribution of "zsize" and if the number of pages with zsize>PAGE_SIZE/2 greatly exceeds the number of pages with "complementary zsize", the frontswap code in zcache can reject the larger pages until balance/sanity is restored.
Quoted 6 lines
Quoted 1 line
Might that help?
We could do that, but I imagine that would let a lot of pages through on most workloads. Ideally, I'd like to find a solution that would capture and (efficiently) store pages that compressed to up to 80% of their original size.
After thinking about this a bit, I have to disagree. For workloads where the vast majority of pages have zsize>PAGE_SIZE/2, this would let a lot of pages through. So if you are correct that LZO is poor at compression and a large majority of pages are in this category, some page-crossing scheme is necessary. However, that isn't what I've seen... the zsize of many swap pages is quite small.
So before commencing on a major compression rewrite, it might be a good idea to measure distribution of zsize for swap pages on a large variety of workloads. This could probably be done by adding a code snippet in the swap path of a normal (non-zcache) kernel. And if the distribution is bad, replacing LZO with a higher-compression-but-slower algorithm might be the best answer, since zcache is replacing VERY slow swap-device reads/writes with reasonably fast compression/decompression. I certainly think that an algorithm approaching an average 50% compression ratio should be the goal.
Dan
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 22 lines
Quoted 7 lines
Quoted 7 lines
One neat feature of frontswap (and the underlying Transcendent Memory definition) is that ANY PUT may be rejected**. So zcache could keep track of the distribution of "zsize" and if the number of pages with zsize>PAGE_SIZE/2 greatly exceeds the number of pages with "complementary zsize", the frontswap code in zcache can reject the larger pages until balance/sanity is restored.
Quoted 6 lines
Quoted 1 line
Might that help?
We could do that, but I imagine that would let a lot of pages through on most workloads. Ideally, I'd like to find a solution that would capture and (efficiently) store pages that compressed to up to 80% of their original size.
After thinking about this a bit, I have to disagree. For workloads where the vast majority of pages have zsize>PAGE_SIZE/2, this would let a lot of pages through. So if you are correct that LZO is poor at compression and a large majority of pages are in this category, some page-crossing scheme is necessary. However, that isn't what I've seen... the zsize of many swap pages is quite small.
Quoted 10 lines
So before commencing on a major compression rewrite, it might be a good idea to measure distribution of zsize for swap pages on a large variety of workloads. This could probably be done by adding a code snippet in the swap path of a normal (non-zcache) kernel. And if the distribution is bad, replacing LZO with a higher-compression-but-slower algorithm might be the best answer, since zcache is replacing VERY slow swap-device reads/writes with reasonably fast compression/decompression. I certainly think that an algorithm approaching an average 50% compression ratio should be the goal.
FWIW, I've measured the distribution of zsize (pages compressed with frontswap) on my favorite workload (kernel "make -j2" on mem=512M to force lots of swapping) and the mean is small, close to 1K (PAGE_SIZE/4). I've added some sysfs shows for both the current and cumulative distribution (0-63 bytes, 64-127 bytes, ..., 4032-4095 bytes) for the next update.
I tried your program on the text of Moby Dick and the mean was still under 1500 bytes ((3*PAGE_SIZE)/8) with a good broad distribution for zsize. I tried your program also on gzip'ed Moby Dick and zcache correctly rejects most of the pages as uncompressible and does fine on other swapped pages.
So I can't reproduce what you are seeing. Somehow you must create and swap a set of pages with a zsize distribution almost entirely between PAGE_SIZE/2 and (PAGE_SIZE*7)/8. How did you do that?
FYI, I also added a sysfs settable for zv_max_page_size... if zsize exceeds it, the page is rejected. It defaults to (PAGE_SIZE*7)/8, which was the non-settable hardwired value before.
Dan
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 10 lines
FWIW, I've measured the distribution of zsize (pages compressed with frontswap) on my favorite workload (kernel "make -j2" on mem=512M to force lots of swapping) and the mean is small, close to 1K (PAGE_SIZE/4). I've added some sysfs shows for both the current and cumulative distribution (0-63 bytes, 64-127 bytes, ..., 4032-4095 bytes) for the next update.
I tried your program on the text of Moby Dick and the mean was still under 1500 bytes ((3*PAGE_SIZE)/8) with a good broad distribution for zsize.
Oops, on retry this morning, I am now clearly seeing the poor compression. Not sure what is different from last night, but I suspect I was "watch"ing the new sysfs output during massive swapping during the run of the test program and it wasn't updated until program completion (at which point I was no longer perusing the sysfs output).
Sorry for the noise. However, now that I have a test case I am implementing another sysfs tunable to reject poorly- compressible pages that would drive the mean zsize above the tunable. Zcache will reject these pages until the mean falls below the threshold. (Setting it to PAGE_SIZE will continue current behavior, but I've set the default to (5*PAGE_SIZE)/8 for now.)
Dan
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Hi Seth,
Quoted 8 lines · Seth Jennings
The real problem here is compressing pages of size x and storing them in a pool that has "chunks", if you will, also of size x, where allocations can't span multiple chunks. Ideally, I'd like to address this issue by expanding the size of the xvmalloc pool chunks from one page to four pages (I can explain why four is a good number, just didn't want to make this note too long).
Quoted 8 lines
After a little playing around, I've found this isn't entirely trivial to do because of the memory mapping implications; more specifically the use of kmap/kunamp in the xvmalloc and zcache layers. I've looked into using vmap to map multiple pages into a linear address space, but it seems like there is a lot of memory overhead in doing that.
Do you have any feedback on this issue or suggestion solution?
xvmalloc fragmentation issue has been reported by several zram users and quite some time back I started working on a new allocator (xcfmalloc) which potentially solves many of these issues. However, all of the details are currently on paper and I'm sure actual implementation will bring a lot of surprises.
Currently, xvmalloc wastes memory due to:
- No compaction support: Each page can store chunks of any size which
makes compaction really hard to implement.
- Use of 0-order pages only: This was enforced to avoid memory
allocation failures. As Dan pointed out, any higher order allocation is almost guaranteed to fail under memory pressure.
To solve these issues, xcfmalloc:
- Supports compaction: Its size class based (like SLAB) which, among
other things, simplifies compaction.
- Supports higher order pages using little trickery:
For 64-bit systems, we can simply use vmalloc(16k or 64k) pages and never bother unmapping them. This is expensive (how much?) in terms of both CPU and memory but easy to implement.
But on 32-bit (almost all "embedded" devices), this ofcourse cannot be done. For this case, the plan is to create a "vpage" abstraction which can be treated as usual higher-order page.
vpage abstraction:
- Allocate 0-order pages and maintain them in an array
- Allow a chunk to cross at most one 4K (or whatever is the nativePAGE_SIZE) page boundary. This limits maximum allocation size to 4K but simplifies mapping logic.
- A vpage is assigned a specific size class just like usual SLAB. This
will simplify compaction.
- xcfmalloc() will return a object handle instead of a direct pointer.
- Provide xcfmalloc_{map,unmap}() which will handle the case where a
chunk spans two pages. It will map the pages using kmap_atomic() and thus user will be expected to unmap them soon.
- Allow vpage to be "partially freed" i.e. empty 4K pages can be freed
individually if completely empty.
Much of this vpage functionality seems to be already present in mainline as "flexible arrays"[1]
For scalability, we can simply go for per-cpu lists and use Hoard[2] like design to bound fragmentation associated with such per-cpu slabs.
Unfortunately, I'm currently too loaded to work on this, atleast for next 2 months (internship) but would be glad to contribute if someone is willing to work on this.
[1] http://lxr.linux.no/linux+v2.6.39/Documentation/flexible-arrays.txt
[2] Hoard allocator: http://www.cs.umass.edu/~emery/pubs/berger-asplos2000.pdf
Thanks,
Nitin
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 2 lines · Nitin Gupta
Much of this vpage functionality seems to be already present in mainline as "flexible arrays"[1]
That's a good observation. I don't know who wrote that junk, but I bet they never thought of using it for this purpose. :)
FWIW, for flex_arrays, the biggest limitation is that the objects currently can not cross page boundaries. The current API also doesn't have any concept of a release function. We'd need those to do the unmapping after a get(). It certainly wouldn't be impossible to fix, but it would probably make it quite a bit more complicated.
The other limitation is that each array can only hold a small number of megabytes worth of data in each array. We only have a single-level table lookup, and that first-level table is limited to PAGE_SIZE (minus a wee bit of metadata).
-- Dave
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 13 lines · Dave Hansen
Quoted 2 lines · Nitin Gupta
Much of this vpage functionality seems to be already present in mainline as "flexible arrays"[1]
That's a good observation. I don't know who wrote that junk, but I bet they never thought of using it for this purpose. :)
FWIW, for
flex_arrays, the biggest limitation is that the objects currently can not cross page boundaries. The current API also doesn't have any concept of a release function. We'd need those to do the unmapping after aget(). It certainly wouldn't be impossible to fix, but it would probably make it quite a bit more complicated.
Quoted 5 lines
The other limitation is that each array can only hold a small number of megabytes worth of data in each array. We only have a single-level table lookup, and that first-level table is limited to
PAGE_SIZE(minus a wee bit of metadata).
These limitations really makes them unsuitable for use in the new allocator and I guess "fixing" them is also not a good idea -- if its fundamentally designed to work on small number of objects, it should probably be left as-is. Not sure who really uses flex arrays?
So, if and when vpage stuff is introduced, existence of flex_arrays should not become a barrier.
Thanks,
Nitin
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Nitin,
I have been working on this xcfmalloc allocator using most of the design points you described (cross-page allocations, compaction). I should have a patch soon.
Quoted 78 lines · Nitin Gupta
Hi Seth,
Quoted 8 lines · Seth Jennings
The real problem here is compressing pages of size x and storing them in a pool that has "chunks", if you will, also of size x, where allocations can't span multiple chunks. Ideally, I'd like to address this issue by expanding the size of the xvmalloc pool chunks from one page to four pages (I can explain why four is a good number, just didn't want to make this note too long).
Quoted 8 lines
After a little playing around, I've found this isn't entirely trivial to do because of the memory mapping implications; more specifically the use of kmap/kunamp in the xvmalloc and zcache layers. I've looked into using vmap to map multiple pages into a linear address space, but it seems like there is a lot of memory overhead in doing that.
Do you have any feedback on this issue or suggestion solution?
xvmalloc fragmentation issue has been reported by several zram users and quite some time back I started working on a new allocator (xcfmalloc) which potentially solves many of these issues. However, all of the details are currently on paper and I'm sure actual implementation will bring a lot of surprises.
Currently, xvmalloc wastes memory due to:
- No compaction support: Each page can store chunks of any size which
makes compaction really hard to implement.
- Use of 0-order pages only: This was enforced to avoid memory
allocation failures. As Dan pointed out, any higher order allocation is almost guaranteed to fail under memory pressure.To solve these issues, xcfmalloc:
- Supports compaction: Its size class based (like SLAB) which, among
other things, simplifies compaction.
- Supports higher order pages using little trickery:For 64-bit systems, we can simply use vmalloc(16k or 64k) pages and never bother unmapping them. This is expensive (how much?) in terms of both CPU and memory but easy to implement.
But on 32-bit (almost all "embedded" devices), this ofcourse cannot be done. For this case, the plan is to create a "vpage" abstraction which can be treated as usual higher-order page.
vpage abstraction:
- Allocate 0-order pages and maintain them in an array
- Allow a chunk to cross at most one 4K (or whatever is the nativePAGE_SIZE) page boundary. This limits maximum allocation size to 4K but simplifies mapping logic.
- A vpage is assigned a specific size class just like usual SLAB. This
will simplify compaction.
-xcfmalloc()will return a object handle instead of a direct pointer.
- Provide xcfmalloc_{map,unmap}() which will handle the case where a
chunk spans two pages. It will map the pages usingkmap_atomic()and thus user will be expected to unmap them soon.
- Allow vpage to be "partially freed" i.e. empty 4K pages can be freed
individually if completely empty.Much of this vpage functionality seems to be already present in mainline as "flexible arrays"[1]
For scalability, we can simply go for per-cpu lists and use Hoard[2] like design to bound fragmentation associated with such per-cpu slabs.
Unfortunately, I'm currently too loaded to work on this, atleast for next 2 months (internship) but would be glad to contribute if someone is willing to work on this.
[1] http://lxr.linux.no/linux+v2.6.39/Documentation/flexible-arrays.txt
[2] Hoard allocator:
http://www.cs.umass.edu/~emery/pubs/berger-asplos2000.pdfThanks,
Nitin
Quoted 7 lines
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>