lkml.org 
[lkml]   [2014]   [May]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2] zram: remove global tb_lock with fine grain lock
On Thu, 15 May 2014 16:00:47 +0800 Weijie Yang <weijie.yang@samsung.com> wrote:

> Currently, we use a rwlock tb_lock to protect concurrent access to
> the whole zram meta table. However, according to the actual access model,
> there is only a small chance for upper user to access the same table[index],
> so the current lock granularity is too big.
>
> The idea of optimization is to change the lock granularity from whole
> meta table to per table entry (table -> table[index]), so that we can
> protect concurrent access to the same table[index], meanwhile allow
> the maximum concurrency.
> With this in mind, several kinds of locks which could be used as a
> per-entry lock were tested and compared:
>
> Test environment:
> x86-64 Intel Core2 Q8400, system memory 4GB, Ubuntu 12.04,
> kernel v3.15.0-rc3 as base, zram with 4 max_comp_streams LZO.
>
> iozone test:
> iozone -t 4 -R -r 16K -s 200M -I +Z
> (1GB zram with ext4 filesystem, take the average of 10 tests, KB/s)
>
> Test base CAS spinlock rwlock bit_spinlock
> -------------------------------------------------------------------
> Initial write 1381094 1425435 1422860 1423075 1421521
> Rewrite 1529479 1641199 1668762 1672855 1654910
> Read 8468009 11324979 11305569 11117273 10997202
> Re-read 8467476 11260914 11248059 11145336 10906486
> Reverse Read 6821393 8106334 8282174 8279195 8109186
> Stride read 7191093 8994306 9153982 8961224 9004434
> Random read 7156353 8957932 9167098 8980465 8940476
> Mixed workload 4172747 5680814 5927825 5489578 5972253
> Random write 1483044 1605588 1594329 1600453 1596010
> Pwrite 1276644 1303108 1311612 1314228 1300960
> Pread 4324337 4632869 4618386 4457870 4500166

Did you investigate seqlocks?

> To enhance the possibility of access the same table[index] concurrently,
> set zram a small disksize(10MB) and let threads run with large loop count.
>
> fio test:
> fio --bs=32k --randrepeat=1 --randseed=100 --refill_buffers
> --scramble_buffers=1 --direct=1 --loops=3000 --numjobs=4
> --filename=/dev/zram0 --name=seq-write --rw=write --stonewall
> --name=seq-read --rw=read --stonewall --name=seq-readwrite
> --rw=rw --stonewall --name=rand-readwrite --rw=randrw --stonewall
> (10MB zram raw block device, take the average of 10 tests, KB/s)
>
> Test base CAS spinlock rwlock bit_spinlock
> -------------------------------------------------------------
> seq-write 933789 999357 1003298 995961 1001958
> seq-read 5634130 6577930 6380861 6243912 6230006
> seq-rw 1405687 1638117 1640256 1633903 1634459
> rand-rw 1386119 1614664 1617211 1609267 1612471
>
> All the optimization methods show a higher performance than the base,
> however, it is hard to say which method is the most appropriate.
>
> On the other hand, zram is mostly used on small embedded system, so we
> don't want to increase any memory footprint.
>
> This patch pick the bit_spinlock method, pack object size and page_flag
> into an unsigned long table.value, so as to not increase any memory
> overhead on both 32-bit and 64-bit system.

bit_spinlocks are not a particularly good or complete mechanism - they
don't have lockdep support and iirc they are somewhat slow.

So we need a pretty good reason to use them. How much memory saving
are we expecting here?

> On the third hand, even though different kinds of locks have different
> performances, we can ignore this difference, because:
> if zram is used as zram swapfile, the swap subsystem can prevent concurrent
> access to the same swapslot;
> if zram is used as zram-blk for set up filesystem on it, the upper filesystem
> and the page cache also prevent concurrent access of the same block mostly.
> So we can ignore the different performances among locks.

So do we need any locking at all?

>
> ....
>
> static void zram_free_page(struct zram *zram, size_t index)
> {
> struct zram_meta *meta = zram->meta;
> unsigned long handle = meta->table[index].handle;
> + int size;
>
> if (unlikely(!handle)) {
> /*
> * No memory is allocated for zero filled pages.
> * Simply clear zero page flag.
> */
> - if (zram_test_flag(meta, index, ZRAM_ZERO)) {
> - zram_clear_flag(meta, index, ZRAM_ZERO);
> + if (zram_test_zero(meta, index)) {
> + zram_clear_zero(meta, index);
> atomic64_dec(&zram->stats.zero_pages);

Having these atomic ops in the alloc/free hotpaths must be costing us?

> }
> return;
>
> ....
>



\
 
 \ /
  Last update: 2014-05-16 00:01    [W:1.234 / U:0.060 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site