Memory Management in NUMA Multicore Systems: Trapped between Cache Contention and Interconnect Overhead

1
Memory Management inNUMA Multicore
SystemsTrapped between Cache Contention and
Interconnect Overhead

• Zoltan Majo and Thomas R. Gross
• Department of Computer Science
• ETH Zurich

2
NUMA multicores
IC
IC
3
NUMA multicores

• Two problems
• NUMA
• interconnect overhead

B
A
IC
IC
MA
MB
4
NUMA multicores

• Two problems
• NUMA
• interconnect overhead
• multicore
• cache contention

B
A
Cache
IC
IC
MA
MB
5
Outline

• NUMA experimental evaluation
• Scheduling
• N-MASS
• N-MASS evaluation

6
Multi-clone experiments

• Memory behavior of unrelated programs

• Intel Xeon E5520
• 4 clones of soplex (SPEC CPU2006)
• local clone
• remote clone

C
C
C
C
C
C
C
C
C
M
M
M
M
C
7
(No Transcript)
8
Performance of schedules

• Which is the best schedule?
• Baseline single-program execution mode

9
Execution time
Slowdown relative to baseline

• local clones
• remote clones
• average

C
C
C
10
Outline

• NUMA experimental evaluation
• Scheduling
• N-MASS
• N-MASS evaluation

11
N-MASS(NUMA-Multicore-Aware Scheduling Scheme)

• Two steps
• Step 1 maximum-local mapping
• Step 2 cache-aware refinement

12
Step 1 Maximum-local mapping
A
MA
B
MB
C
MC
D
MD
13
Default OS scheduling
B
A
D
C
MB
MA
MC
MD
14
N-MASS(NUMA-Multicore-Aware Scheduling Scheme)

• Two steps
• Step 1 maximum-local mapping
• Step 2 cache-aware refinement

15
Step 2 Cache-aware refinement

• In an SMP

D
C
MB
MA
MC
MD
16
Step 2 Cache-aware refinement

• In an SMP

B
A
D
C
MB
MA
MC
MD
MA
17
Step 2 Cache-aware refinement

• In an SMP

A
B
C
D
18
Step 2 Cache-aware refinement

• In a NUMA

B
A
C
D
MB
MA
MC
MD
19
Step 2 Cache-aware refinement

• In a NUMA

A
D
C
B
MB
MA
MC
MD
20
Step 2 Cache-aware refinement

• In a NUMA

A
B
C
D
21
Performance factors

• Two factors cause performance degradation
• NUMA penalty
• slowdown due toremote memory access
• cache pressure
• local processesmisses / KINST (MPKI)
• remote processesMPKI x NUMA penalty

22
Implementation

• User-mode extension to the Linux scheduler
• Performance metrics
• hardware performance counter feedback
• NUMA penalty
• perfect information from program traces
• estimate based on MPKI
• All memory for a process allocated on one
  processor

23
Outline

• NUMA experimental evaluation
• Scheduling
• N-MASS
• N-MASS evaluation

24
Workloads
NUMA penalty

• SPEC CPU2006 subset
• 11 multi-program workloads (WL1 ? WL11)
• 4-program workloads(WL1 ? WL9)
• 8-program workloads(WL10, WL11)

CPU-bound
Memory-bound
25
Memory allocation setup

• Where the memory of each process is allocated
  influences performance
• Controlled setup memory allocation maps

26
Memory allocation maps
B
MB
A
C
MC
D
MD
M A
Allocation map 0000
27
Memory allocation maps
B
A
C
D
28
Memory allocation maps
B
A
C
D
Unbalanced
Balanced
29
Evaluation

• Baseline Linux average
• Linux scheduler non-deterministic
• average performance degradation in all possible
  cases
• N-MASS with perfect NUMA penalty information

30
WL9 Linux average
Average slowdown relative to single-program mode
31
WL9 N-MASS
Average slowdown relative to single-program mode
32
WL1 Linux average and N-MASS
Average slowdown relative to single-program mode
33
N-MASS performance

• N-MASS reduces performance degradation by up to
  22
• Which factor more important interconnect
  overhead or cache contention?
• Compare
• - maximum-local
• - N-MASS (maximum-local cache
  refinement step)

34
Data-locality vs. cache balancing (WL9)
Performance improvement relative to Linux average
35
Data-locality vs. cache balancing (WL1)
Performance improvement relative to Linux average
36
Data locality vs. cache balancing

• Data-locality more important than cache
  balancing
• Cache-balancing gives performance benefits
  mostly with unbalanced allocation maps
• What if information about NUMA penalty not
  available?

37
Estimating NUMA penalty
NUMA penalty

• NUMA penalty is not directly measurable
• Estimate fit linear regression onto MPKI data

38
Estimate-based N-MASS performance
Performance improvement relative to Linux average
39
Conclusions

• N-MASS NUMA?multicore-aware scheduler
• Data locality optimizations more beneficial than
  cache contention avoidance
• Better performance metrics needed for scheduling

40
Thank you! Questions?