Enable system hashtable memory to be distributed among nodes on x86_64 NUMA
Forcing the kernel to use node interleaved vmalloc instead of bootmem for
the system hashtable memory (alloc_large_system_hash) reduces the memory
imbalance on node 0 by around 40MB on a 8 node x86_64 NUMA box:
Before the following patch, on bootup of a 8 node box:
Node 0 MemTotal: 3407488 kB
Node 0 MemFree: 3206296 kB
Node 0 MemUsed: 201192 kB
Node 0 Active: 7012 kB
Node 0 Inactive: 512 kB
Node 0 Dirty: 0 kB
Node 0 Writeback: 0 kB
Node 0 FilePages: 1912 kB
Node 0 Mapped: 420 kB
Node 0 AnonPages: 5612 kB
Node 0 PageTables: 468 kB
Node 0 NFS_Unstable: 0 kB
Node 0 Bounce: 0 kB
Node 0 Slab: 5408 kB
Node 0 SReclaimable: 644 kB
Node 0 SUnreclaim: 4764 kB
After the patch (or using hashdist=1 on the kernel command line):
Node 0 MemTotal: 3407488 kB
Node 0 MemFree: 3247608 kB
Node 0 MemUsed: 159880 kB
Node 0 Active: 3012 kB
Node 0 Inactive: 616 kB
Node 0 Dirty: 0 kB
Node 0 Writeback: 0 kB
Node 0 FilePages: 2424 kB
Node 0 Mapped: 380 kB
Node 0 AnonPages: 1200 kB
Node 0 PageTables: 396 kB
Node 0 NFS_Unstable: 0 kB
Node 0 Bounce: 0 kB
Node 0 Slab: 6304 kB
Node 0 SReclaimable: 1596 kB
Node 0 SUnreclaim: 4708 kB
I guess it is a good idea to keep HASHDIST_DEFAULT "on" for x86_64 NUMA
since x86_64 has no dearth of vmalloc space? Or maybe enable hash
distribution for all 64bit NUMA arches? The following patch does it only
for x86_64.
I ran a HPC MPI benchmark -- 'Ansys wingsolid', which takes up quite a bit of
memory and uses up tlb entries. This was on a 4 way, 2 socket
Tyan AMD box (non vsmp), with 8G total memory (4G pernode).
The results with and without hash distribution are:
1. Vanilla - runtime of 1188.000s
2. With hashdist=1 runtime of 1154.000s
Oprofile output for the duration of run is:
1. Vanilla:
PU: AMD64 processors, speed 2411.16 MHz (estimated)
Counted L1_AND_L2_DTLB_MISSES events (L1 and L2 DTLB misses) with a unit
mask of 0x00 (No unit mask) count 500
samples % app name symbol name
163054 6.5513 libansys1.so MultiFront::decompose(int, int,
Elemset *, int *, int, int, int)
162061 6.5114 libansys3.so blockSaxpy6L_fd
162042 6.5107 libansys3.so blockInnerProduct6L_fd
156286 6.2794 libansys3.so maxb33_
87879 3.5309 libansys1.so elmatrixmultpcg_
84857 3.4095 libansys4.so saxpy_pcg
58637 2.3560 libansys4.so .st4560
46612 1.8728 libansys4.so .st4282
43043 1.7294 vmlinux-t copy_user_generic_string
41326 1.6604 libansys3.so blockSaxpyBackSolve6L_fd
41288 1.6589 libansys3.so blockInnerProductBackSolve6L_fd
2. With hashdist=1
CPU: AMD64 processors, speed 2411.13 MHz (estimated)
Counted L1_AND_L2_DTLB_MISSES events (L1 and L2 DTLB misses) with a unit
mask of 0x00 (No unit mask) count 500
samples % app name symbol name
162993 6.9814 libansys1.so MultiFront::decompose(int, int,
Elemset *, int *, int, int, int)
160799 6.8874 libansys3.so blockInnerProduct6L_fd
160459 6.8729 libansys3.so blockSaxpy6L_fd
156018 6.6826 libansys3.so maxb33_
84700 3.6279 libansys4.so saxpy_pcg
83434 3.5737 libansys1.so elmatrixmultpcg_
58074 2.4875 libansys4.so .st4560
46000 1.9703 libansys4.so .st4282
41166 1.7632 libansys3.so blockSaxpyBackSolve6L_fd
41033 1.7575 libansys3.so blockInnerProductBackSolve6L_fd
35762 1.5318 libansys1.so inner_product_sub
35591 1.5245 libansys1.so inner_product_sub2
28259 1.2104 libansys4.so addVectors
Signed-off-by: Pravin B. Shelar <pravin.shelar@calsoftinc.com>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Christoph Lameter <clameter@engr.sgi.com>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
