Adaptive page grouping reduces energy consumption in hybrid PRAM-DRAM main memory systems in 3 ways: 1) It groups pages with similar access patterns together, lowering the number of writes to the higher-power PRAM pages. 2) By migrating hot pages to DRAM and cold pages to PRAM based on access patterns, it reduces the amount of data migration between the two memories. 3) This targeted migration approach maintains comparable memory system performance to other state-of-the-art hybrid memory management techniques while lowering overall energy usage by an average of 24%.

1. Adaptive Page Grouping for Energy Efficiency
in Hybrid PRAM-DRAM Main Memory
Dong-Jae Shin, Sung Kyu Park,
Seong Min Kim and Kyu Ho Park
Korea Advanced Institute of Science & Technology
Presenter : Dong-Jae Shin
Date : 2012/10/24

2. OUTLINE
1 Introduction
2 Motivation
3 Adaptive Page Grouping System
4 Evaluation
5 Conclusion
2 / 22

3. 1. Introduction
Energy Consumption in Computer Systems
• Energy consumption of computer system is increasing
– Data centers consumes 1.5% of the world's energy use and doubled
in 5 years [1]
– DRAM is significant factor for consumed energy
 Effects
Networking
5%  Electric charge is more than 10%
of operation cost in data centers[1]
DRAM
30%  It continues to increase steadily
CPUs
33%  Challenging issues
 How can we reduce energy
consumption in main memory
 Consumed power in memory
Other Disks
22% 10%  Static power
 Dynamic power
Distribution of power usage by Google Data Center [2]
[1] Harnessing Green IT: Principles and Practices,” IEEE IT Professional, January–February 2008
[2] L. Barroso and U. Holzle. The Datacenter as a Computer: An Introduction to the Design of 3 / 22
Warehouse-Scale Machines. Morgan& Claypool, 2009

4. 1. Introduction
DRAM and Phase change memory(PRAM)
• DRAM • PRAM
– Charge capacitor and read – Store value with resistance
– Leakage current – No refresh power
– Need refresh power – Non-volatile
– Limit of capacitor size – Multi-level Cell(MLC)[4] available
DRAM Cell [3] PRAM Cell [3]
(Charge Memory) (Resistive Memory)
 Good performance  Low static power
 High static power  Relatively low performance
[3] Architecting Phase Change Memory as a Scalable DRAM Alternative, ISCA 2009
[4] Drift-Tolerant Multilevel Phase-Change Memory, IMW 2011 4 / 22

5. 1. Introduction
DRAM & PRAM Technology
• Single Chip Capacity DRAM PRAM
Design Rule 8Gb
4Gb
20nm 2Gb 8Gb
40nm
60nm 1Gb
512Mb
80nm
256Mb
100nm 256Mb
Year
2004 2006 2008 2010 2012
[5] Memory Technology and Solutions Roadmap, SAMSUNG
5 / 22

6. 2. Motivation
PRAM as a main memory
Summarized comparison between DRAM and PRAM [6]
Attributes DRAM PRAM
Non-volatility No Yes
Read latency 10~60ns 48ns
Write latency 10~60ns 40~150ns
Read energy 2pJ / bit 2pJ / bit
Write energy 2pJ / bit 10pJ / bit
Idle power ~W/GB ~0.05W.GB
Write Endurance 1015 108
• Challenging issues of
PRAM main memory
– Endurance problem Only PRAM is
• Less than 1 years not solution!
– High write latency
– High write power
[6] Non-Volatile Memory: Emerging Technologies And Their Impacts on Memory Systems, 2010
6 / 22

7. 2. Motivation
Hybrid Main Memory Architecture
CPU
• Motivation
– To exploit both of
Core Core
high performance of DRAM and low
Core Core static power of PRAM
Last Level Cache
• Necessity for management
Memory Controller
– In worst case (Hot pages in PRAM)
• consumes high dynamic power
• has high latency
Linear address space
• PRAM cell wears out rapidly
DRAM PRAM
Hybrid Main Memory Architecture
 Direct access to PRAM  Elaborate management is needed
7 / 22

8. 2. Motivation
Access Pattern with Spatial Locality
• Access count and difference of PFN(Page Frame Number)
gzip mcf
Spatial locality was found over pages
8 / 22

9. 3. GPM System
Overall System Architecture
• Hybrid memory management daemon works periodically
– Monitoring module
• Look up and store access information of pages for hot/cold decision
– Page Grouping Module
• Adaptive page grouping
– Migration Module
• Page placement DRAM
Access Page
Kernel Table
Hybrid Memory
Management Daemon Hot Page
Group
Monitoring & Migration
Shifting Module
Migration
Page Grouping Cold Page
Module Group
Migration
Migration Module
(DRAM  PRAM)
PRAM
Free Page Hot Group Cold Group
9 / 22

10. 3. GPM System
4-1. Store Writing History
Hybrid Memory
Management Daemon
• Dirty bit shifting
Monitoring &
– Store writing history per page for hot/cold decision
Shifting Module
– Use available space in a page table entry (PTE)
Page Grouping • No additional space for storing history, different from
Module
other works
Migration Module
Page Table Entry D = Dirty bit
(DRAM  PRAM)
63 62 61 60 59 52 51 12 11 9 8 7 6 5 4 3 2 1 0
P P P U R
N Used by
4 3 2 1 Available Physical-Page Frame Number G A D A C W / / P
X OS
T D T S W
To store
dirty bit
4th 3rd 2nd 1st Dirty Bit
Dirty Bit
v v v Monitoring
v v and
Page Table Bit Shifting
(512 entries v v v v
per directory)
…
Bit Shifting v : bit is 1 v : bit will be 1 10 / 22

11. 3. GPM System
4-2. Adaptive Page Grouping
Hybrid Memory • Adaptive page grouping
Management Daemon – Physically near pages have similar access count due to Linux
Monitoring & buddy system
– These pages are grouped  Migration ↓, Accuracy↑
Shifting Module
Page Grouping
Module
– Algorithm
• Step ① : Calculate difference between Page Frame Numbers(PFN)
Migration Module
(DRAM  PRAM)
• Step ② : If (diff. < threshold), then add to group
① ②
63 62 61 60 59 52 51 12 11 9 8 7 6 5 4 3 2 1 0
Diff. of
P P P U R PFN Grouping
N Used by PFN
4 3 2 1 Available Physical-Page Frame Number G A D A C W / / P
X OS
T D T S W 0x000C0
Page Table Entry 0x000C1 1
Group
0x000C2 1
0x000C3 1
Linux Buddy System
0x02041 8062
4KB
0x02042 1 Group
Memory Block
8KB 0x02045 3
16KB 0x08892 26701 X
0x0631B 9591 X
…
Page Table 11 / 22

12. 3. GPM System
4-3. Hot/Cold Criterion
Hybrid Memory • Hot cold decision for migration
Management Daemon
– Average of hotness in page group
– Weighted value to exploit temporal locality
Monitoring &
Shifting Module
Page Grouping
63 62 61 60 59 52 51
Page Table Entry 12 11 9 8 7 6 5 4 3 2 1 0
Module
P P P U R
N Used by
4 3 2 1 Available Physical-Page Frame Number G A D A C W / / P
Migration Module X OS
T D T S W
(DRAM  PRAM)
① ② ③ ④
4th 3rd 2nd 1st Hotness Avg.
(Weight 1) (Weight 2) (Weight 4) (Weight 8) Decision Characteristic Criterion
(page) Hotness
v v 9
Hot Avg. >= Hot threshold
v v 12
11.25 Hot
v v 10
Page Table
v v v 14 Hot threshold >
Warm
Avg. > Cold threshold
v 1
v 1 1.33 Cold
v 2 Cold Cold threshold >= Avg.
v v 6 6 Warm
v v 9 9 Warm
12 / 22

13. 3. GPM System
4-4. Migration Policy
Hybrid Memory • First allocation from DRAM
Management Daemon
– Firstly allocated page is used immediately[7]
Monitoring &
Shifting Module
• Kernel memory is allocated to DRAM
– Kernel memory is frequently used[7]
• Hot groups to DRAM, and Cold groups to PRAM
Page Grouping
Module
Migration Module
– To exploit both advantages of DRAM and PRAM
(DRAM  PRAM)
• Warm groups do not migrate
– To reduce migration
4th 3rd 2nd 1st Hotness Avg.
(Weight 1) (Weight 2) (Weight 4) (Weight 8) Decision
(page) Hotness
v v 9
If page group are in
v v 12
11.25 Hot PRAM,
v v 10 migrate to DRAM
Page Table v v v 14
v 1
If pages group are in
v 1 1.33 Cold DRAM,
v 2 migrate to PRAM
v v 6 6 Warm
Do not migrate
v v 9 9 Warm
[7] Page Placement in Hybrid Memory Systems, ICS 2011 13 / 22

14. 4. Evaluation
Experiment Environment
• Implementation Hybrid Memory
– Implement in Linux kernel as a Management Daemon
kernel thread (daemon) Monitoring &
– without additional hardware
Shifting Module
 only software patch Page Grouping
Module
• Specification
Migration Module
(DRAM  PRAM)
– CPU : Intel Dual Core 2.4Ghz
– Memory : 1GB DRAM, 1GB PRAM
– OS : Linux 2.6.31 kernel
• Power & Timing calculation
– Counting access count using Pin-tool[8]
[8] Pin: building customized program analysis tools with dynamic instrumentation, PLDI 2005
14 / 22

15. 4. Evaluation
Workload and Parameters
• Workload ( SPEC 2000 + MaaS[9] ) • DRAM & PRAM characteristic[3]
Workload Working Set Size Parameter DRAM PRAM
Power characteristics (per GB)
176.gcc.ref 97MB
Row read power 210mW 78mW
188.ammp.ref 40MB
Row write power 195mW 773mW
172.mgrid.ref 69MB Active power 75mW 25mW
183.equake.ref 64MB Standby power 90mW 45mW
181.mcf.ref 165MB Refresh power 4mW 0mW
Timing characteristics
164.gzip.ref 193MB
Initial row read 15ns 28ns
173.applu.ref 195MB Initial row write 22ns 150ns
MaaS 480MB Same row read/write 15ns 15ns
• Comparable System
– DRAM only (2GB) , PRAM only (2GB)
– Hybrid memory (DRAM 1GB , PRAM 1GB)
• PDRAM [10] , Second chance algorithm[11], Adaptive Page Grouping (APG)
[9] Multimedia Matching as a Service: Technical Challenges and Blueprints, ITC-CSCC 2011
[10] PDRAM : A Hybrid PRAM and DRAM Main Memory System, DAC 2009
[11] Power-Aware Memory Management for Hybrid Main Memory, ICNIT 2011 15 / 22

16. 4. Evaluation
Energy Consumption of Memory System
Normalized Energy Consumption in Memory System
DRAM Only PRAM Only PDRAM S.chance APG
2
Normalized to DRAM
1.5
1
0.5
0
mcf gzip applu gcc equake ammp mgrid maas avg.
• 24% of energy was reduced compared to DRAM
• 8.5% was reduced compared to power-aware management[7]
16 / 22

17. 4. Evaluation
Reason 1 : Reduction of Write count in PRAM
Write Count per PRAM Page
1500
PDRAM
S.chance
APG
1000
Write Count
500
0
mcf gzip applu gcc equake ammp mgrid maas avg.
Write count was effectively reduced by Adaptive Page Grouping
17 / 22

18. 4. Evaluation
Reason 2 : Reduction of Migration Energy
Total Migration Energy
1200
PRAM->DRAM
1000 DRAM->PRAM
Migration Energy (J)
800
600
400
200
0
APG
APG
APG
APG
APG
APG
APG
APG
APG
S.chance
S.chance
S.chance
S.chance
S.chance
S.chance
S.chance
S.chance
S.chance
On average, 76% of migration energy was reduced compared to
power-aware management[11]
18 / 22

19. 4. Evaluation
Performance of Memory System
Normalized Access Time
DRAM Only PRAM Only PDRAM S.chance APG
5
4
Normalized to DRAM
3
2
1
0
mcf gzip applu gcc equake ammp mgrid maas avg.
On average, only 8% of access was increased compared to DRAM,
38% was reduced compared to power-aware management[11]
19 / 22

20. 5. Conclusion
Overhead Analysis
• Overhead factor • Overhead result
– Algorithm overhead – Including 4 factors
• Page table scanning Overhead
Workload
• Bit shifting (%)
• Adaptive page grouping mcf 0.63
– Migration overhead gzip 0.52
applu 0.41
gcc 0.88
equake 0.32
ammp 0.15
mgrid 0.22
MaaS 1.23
On average, less than 1% of overhead
average 0.55
20 / 22

21. 5. Conclusion
Conclusion
• Design and implement Adaptive Page Grouping
– Dirty bit shifting
• Storing access information per page without additional space
• Weighted dirty bit for exploiting temporal locality
– Adaptive page grouping
• Exploit spatial locality over pages
• Reduce migration operation between DRAM and PRAM
– Grouped page migration
• Effectively reduce write operation on PRAM
• Power saving
– 24% reduction compared to DRAM only system
– 8.5% reduction compared to power aware management[7]
– With less than 1% performance degradation
Granularity Locality Level Additional H/W
Temporal
Adaptive Page Grouping Page Group OS none
& Spatial
[7] Power-Aware Memory Management for Hybrid Main Memory, ICNIT 2011 21 / 22

22. 5. Conclusion
Further Works
• Further works
– Evaluation on mobile system
• Power saving increases lifetime of battery
– Combining with Wear leveling scheme
• Word-level shifting
• Line shifting
• Page swapping
22 / 22

23. 6. Conclusion
Reference
• [1] Harnessing Green IT: Principles and Practices,” IEEE IT Professional, January–
February 2008
• [2] L. Barroso and U. Holzle. The Datacenter as a Computer: An Introduction to the
Design of Warehouse-Scale Machines. Morgan& Claypool, 2009
• [3] Architecting Phase Change Memory as a Scalable DRAM Alternative, ISCA 2009
• [4] Drift-Tolerant Multilevel Phase-Change Memory, IMW 2011
• [5] Memory Technology and Solutions Roadmap, SAMSUNG
• [6] Non-Volatile Memory: Emerging Technologies And Their Impacts on Memory
Systems, 2010
• [7] Page Placement in Hybrid Memory Systems, ICS 2011
• [8] Pin: building customized program analysis tools with dynamic
instrumentation, PLDI 2005
• [9] Multimedia Matching as a Service: Technical Challenges and Blueprints, ITC-
CSCC 2011
• [10] PDRAM : A Hybrid PRAM and DRAM Main Memory System, DAC 2009
• [11] Power-Aware Memory Management for Hybrid Main Memory, ICNIT 2011
23 / 22

24. Thank you