From: Dmitry Eremin <dmitry.ere...@intel.com>
The main idea to have more CPU partitions is based on KNL experience.
When a thread submit IO for network communication one of threads from
current CPT is used for network stack. Whith high parallelization many
threads become involved in network submission but having less CPU
partitions they will wait until single thread process them from network
queue. So, the bottleneck just moves into network layer in case of
small amount of CPU partitions. My experiments showed that the best
performance was when for each IO thread we have one network thread.
This condition can be provided having 2 real HW cores (without hyper
threads) per CPT. This is exactly what implemented in this patch.
Change CPT estimate algorithm from 2 * (N - 1)^2 < NCPUS <= 2 * N^2
to 2 HW cores per CPT. This is critical for machines with number of
cores different from 2^N.
Current algorithm splits CPTs in KNL:
LNet: HW CPU cores: 272, npartitions: 16
cpu_partition_table=
0 : 0-4,68-71,136-139,204-207
1 : 5-9,73-76,141-144,209-212
2 : 10-14,78-81,146-149,214-217
3 : 15-17,72,77,83-85,140,145,151-153,208,219-221
4 : 18-21,82,86-88,150,154-156,213,218,222-224
5 : 22-26,90-93,158-161,226-229
6 : 27-31,95-98,163-166,231-234
7 : 32-35,89,100-103,168-171,236-239
8 : 36-38,94,99,104-105,157,162,167,172-173,225,230,235,240-241
9 : 39-43,107-110,175-178,243-246
10 : 44-48,112-115,180-183,248-251
11 : 49-51,106,111,117-119,174,179,185-187,242,253-255
12 : 52-55,116,120-122,184,188-190,247,252,256-258
13 : 56-60,124-127,192-195,260-263
14 : 61-65,129-132,197-200,265-268
15 : 66-67,123,128,133-135,191,196,201-203,259,264,269-271
New algorithm will split CPTs in KNL:
LNet: HW CPU cores: 272, npartitions: 34
cpu_partition_table=
0 : 0-1,68-69,136-137,204-205
1 : 2-3,70-71,138-139,206-207
2 : 4-5,72-73,140-141,208-209
3 : 6-7,74-75,142-143,210-211
4 : 8-9,76-77,144-145,212-213
5 : 10-11,78-79,146-147,214-215
6 : 12-13,80-81,148-149,216-217
7 : 14-15,82-83,150-151,218-219
8 : 16-17,84-85,152-153,220-221
9 : 18-19,86-87,154-155,222-223
10 : 20-21,88-89,156-157,224-225
11 : 22-23,90-91,158-159,226-227
12 : 24-25,92-93,160-161,228-229
13 : 26-27,94-95,162-163,230-231
14 : 28-29,96-97,164-165,232-233
15 : 30-31,98-99,166-167,234-235
16 : 32-33,100-101,168-169,236-237
17 : 34-35,102-103,170-171,238-239
18 : 36-37,104-105,172-173,240-241
19 : 38-39,106-107,174-175,242-243
20 : 40-41,108-109,176-177,244-245
21 : 42-43,110-111,178-179,246-247
22 : 44-45,112-113,180-181,248-249
23 : 46-47,114-115,182-183,250-251
24 : 48-49,116-117,184-185,252-253
25 : 50-51,118-119,186-187,254-255
26 : 52-53,120-121,188-189,256-257
27 : 54-55,122-123,190-191,258-259
28 : 56-57,124-125,192-193,260-261
29 : 58-59,126-127,194-195,262-263
30 : 60-61,128-129,196-197,264-265
31 : 62-63,130-131,198-199,266-267
32 : 64-65,132-133,200-201,268-269
33 : 66-67,134-135,202-203,270-271
'N' pattern in KNL works is not always good.
in flat mode it will be one CPT with all CPUs inside.
in SNC-4 mode:
cpu_partition_table=
0 : 0-17,68-85,136-153,204-221
1 : 18-35,86-103,154-171,222-239
2 : 36-51,104-119,172-187,240-255
3 : 52-67,120-135,188-203,256-271
Signed-off-by: Dmitry Eremin <dmitry.ere...@intel.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8703
Reviewed-on: https://review.whamcloud.com/24304
Reviewed-by: James Simmons <uja.o...@yahoo.com>
Reviewed-by: Andreas Dilger <andreas.dil...@intel.com>
Reviewed-by: Oleg Drokin <oleg.dro...@intel.com>
Signed-off-by: James Simmons <jsimm...@infradead.org>
---
.../staging/lustre/lnet/libcfs/linux/linux-cpu.c | 30 ++++------------------
1 file changed, 5 insertions(+), 25 deletions(-)
diff --git a/drivers/staging/lustre/lnet/libcfs/linux/linux-cpu.c
b/drivers/staging/lustre/lnet/libcfs/linux/linux-cpu.c
index 915cfca..ae5fd16 100644
--- a/drivers/staging/lustre/lnet/libcfs/linux/linux-cpu.c
+++ b/drivers/staging/lustre/lnet/libcfs/linux/linux-cpu.c
@@ -768,34 +768,14 @@ static int cfs_cpt_choose_ncpus(struct cfs_cpt_table
*cptab, int cpt,
static int cfs_cpt_num_estimate(void)
{
- int nnode = num_online_nodes();
+ int nthr = cpumask_weight(topology_sibling_cpumask(smp_processor_id()));
int ncpu = num_online_cpus();
- int ncpt;
+ int ncpt = 1;
- if (ncpu <= CPT_WEIGHT_MIN) {
- ncpt = 1;
- goto out;
- }
-
- /* generate reasonable number of CPU partitions based on total number
- * of CPUs, Preferred N should be power2 and match this condition:
- * 2 * (N - 1)^2 < NCPUS <= 2 * N^2
- */
- for (ncpt = 2; ncpu > 2 * ncpt * ncpt; ncpt <<= 1)
- ;
-
- if (ncpt <= nnode) { /* fat numa system */
- while (nnode > ncpt)
- nnode >>= 1;
+ if (ncpu > CPT_WEIGHT_MIN)
+ for (ncpt = 2; ncpu > 2 * nthr * ncpt; ncpt++)
+ ; /* nothing */
- } else { /* ncpt > nnode */
- while ((nnode << 1) <= ncpt)
- nnode <<= 1;
- }
-
- ncpt = nnode;
-
-out:
#if (BITS_PER_LONG == 32)
/* config many CPU partitions on 32-bit system could consume
* too much memory
--
1.8.3.1
