Re: [PATCH v8 14/14] mm: zswap: Compress batching with request chaining in zswap_store() of large folios.

[Yosry Ahmed]

On Mon, Mar 03, 2025 at 09:34:04PM +0000, Sridhar, Kanchana P wrote:
> 
> > -----Original Message-----
> > From: Nhat Pham <nphamcs@xxxxxxxxx>
> > Sent: Monday, March 3, 2025 10:22 AM
> > To: lkp <lkp@xxxxxxxxx>
> > Cc: Sridhar, Kanchana P <kanchana.p.sridhar@xxxxxxxxx>; linux-
> > kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; hannes@xxxxxxxxxxx;
> > yosry.ahmed@xxxxxxxxx; chengming.zhou@xxxxxxxxx;
> > usamaarif642@xxxxxxxxx; ryan.roberts@xxxxxxx; 21cnbao@xxxxxxxxx;
> > ying.huang@xxxxxxxxxxxxxxxxx; akpm@xxxxxxxxxxxxxxxxxxxx; linux-
> > crypto@xxxxxxxxxxxxxxx; herbert@xxxxxxxxxxxxxxxxxxx;
> > davem@xxxxxxxxxxxxx; clabbe@xxxxxxxxxxxx; ardb@xxxxxxxxxx;
> > ebiggers@xxxxxxxxxx; surenb@xxxxxxxxxx; Accardi, Kristen C
> > <kristen.c.accardi@xxxxxxxxx>; llvm@xxxxxxxxxxxxxxx; oe-kbuild-
> > all@xxxxxxxxxxxxxxx; Feghali, Wajdi K <wajdi.k.feghali@xxxxxxxxx>; Gopal,
> > Vinodh <vinodh.gopal@xxxxxxxxx>
> > Subject: Re: [PATCH v8 14/14] mm: zswap: Compress batching with request
> > chaining in zswap_store() of large folios.
> > 
> > On Mon, Mar 3, 2025 at 3:07 AM kernel test robot <lkp@xxxxxxxxx> wrote:
> > >
> > > Hi Kanchana,
> > >
> > > kernel test robot noticed the following build errors:
> > >
> > > > 1166                          prefetchw(entries[j]);
> > > --
> > 
> > Why are we doing this anyway? Does it have a notable performance
> > difference? At the very least, leave a comment explaining why we're
> > prefetching this (although the build error suggests that we have to
> > remove it anyway).
> 
> Hi Nhat,
> 
> Yes, it does. The use of prefetchw reduces sys time by ~1.5% because
> it minimizes cache-miss latency by moving the zswap entry to the cache
> before it is written to. 
> 
> This is data with kernel compilation test, v8 without prefetchw and v8 as-is:
> 
> --------------------------------------------------------------------------------
>  Kernel compile       v8 without               v8      v8 without              v8
>  allmodconfig          prefetchw                        prefetchw
>  2M folios
>  --------------------------------------------------------------------------------
>  zswap compressor    deflate-iaa      deflate-iaa            zstd            zstd   
>  --------------------------------------------------------------------------------
>  real_sec                 732.89           735.63          768.53          758.21
>  user_sec              15,708.37        15,699.84       15,702.64       15,678.73
>  sys_sec                4,632.58         4,563.70        5,735.06        5,635.69
>  --------------------------------------------------------------------------------
>  Max_Res_Set_Size_KB   1,874,672        1,867,516       1,874,684       1,872,888
>  --------------------------------------------------------------------------------
>  memcg_high                    0                0               0               0
>  memcg_swap_fail               0                0               0               0
>  zswpout             114,742,930      112,836,725      92,904,961      89,596,085
>  zswpin               41,184,897       39,983,793      31,018,149      29,163,932
>  pswpout                     625            1,069             558           1,059
>  pswpin                      599            1,056             540           1,051
>  thp_swpout                    1                2               1               2
>  thp_swpout_fallback      10,967           10,195           6,918           6,141
>  pgmajfault           42,588,331       41,349,069      31,931,882      30,006,422
>  ZSWPOUT-2048kB            7,661            8,710           6,799           7,480
>  SWPOUT-2048kB                 1                2               1               2
>  --------------------------------------------------------------------------------
> 
> 
> Sure, I will add a comment, and also "#include <linux/prefetch.h>" in zswap.c
> that will resolve the build error. This is similar to how these files handle prefetchw:
> mm/vmscan.c, kernel/locking/qspinlock.c, include/asm-generic/xor.h, etc.

Please also explicitly mention that the prefetch and likely/unlikely
annotations prevent regressions with software compression like zstd, and
generally improve the performance with the batching code by ~1.5%.

> 
> Thanks,
> Kanchana
>