SLIDE 1
2
Kingsguard: Write-Rationing Garbage Collection for Hybrid Memories - - PowerPoint PPT Presentation
Kingsguard: Write-Rationing Garbage Collection for Hybrid Memories - - PowerPoint PPT Presentation
Kingsguard: Write-Rationing Garbage Collection for Hybrid Memories Shoaib Akram (Ghent) , Jennifer B. Sartor (Ghent), Kathryn S. Mckinley (Google), and Lieven Eeckhout (Ghent) Shoaib.Akram@UGent.be DRAM is facing challenges Scalability
SLIDE 2
SLIDE 3
Phase change memory is promising
3
GB/$ J Latency L Endurance L
3
temperature set to crystalline read reset to amorphous time
But …
SLIDE 4
4
Hybrid DRAM-PCM memory
Challenge Mitigate PCM wear-out and extend its lifetime Speed Endurance Energy Capacity
DRAM PCM
SLIDE 5
5
Wear Level Wear Level Wear Level Operating System
Language runtime
How to mitigate PCM wear-out?
Phase change memory as …
SLIDE 6
6
1 2 Lifetime in years
32 GB PCM memory, 32 cores
PCM only with leveling is not practical
SLIDE 7
7
OS to limit PCM writes
Drawbacks Coarse grained Page migrations can be costly
DRAM PCM
SLIDE 8
8
Managed runtime to limit PCM writes
Our work uses garbage collection to keep highly written objects in DRAM
nursery
mature
- bserver
PCM DRAM
mature
SLIDE 9
nursery mature
GC
70%
- f writes
9
Distribution of writes in GC runtime
SLIDE 10
nursery mature
GC
22%
to 2% of objects
70%
- f writes
10
Distribution of writes in GC runtime
SLIDE 11
mature
GC
Write-Rationing Garbage Collectors Contribution DRAM PCM
11
SLIDE 12
Kingsguard- Nursery Kingsguard- Writers
12
Two write-rationing garbage collectors
SLIDE 13
13
Heap organization in DRAM
nursery
mature large GC DRAM
SLIDE 14
14
mature large GC DRAM PCM
KG-N Kingsguard-Nursery
nursery
SLIDE 15
15
KG-W Kingsguard-Writers
mature large
- bserver
PCM
mature large
DRAM
nursery
SLIDE 16
16
Observing writes
Write barrier configurations Observe references Observe references and primitives Write barrier sets a header bit on object writes references primitives header Object format
SLIDE 17
17
Additional optimizations in KG-W
Large object optimization Allocate selected large objects in DRAM Metadata optimization Allocate PCM metadata in DRAM
SLIDE 18
nursery
½ of remaining nursery
large
Monitor PCM write rate to turn opt on/off
18
Large object optimization
SLIDE 19
19
Metadata optimization
Mature Meta Full-heap GC: Mark live PCM objects KG-W: Keep mark bytes of PCM objects in DRAM
SLIDE 20
20
Metadata optimization
Mature Meta Full-heap GC: Mark live PCM objects KG-W: Keep mark bytes of PCM objects in DRAM address_mark_bit = start_meta + idx_pcm_obj
SLIDE 21
21
Evaluation Methodology
(1) Simulator (2) Real Java applications Jikes research virtual machine
Hardware Software
SLIDE 22
22
Simulation with Sniper
7 DaCapo applications 4 cores, 1 MB per core LLC Scale simulated rates to a 32 core machine using trends from real hw
SLIDE 23
Memory systems
Homogeneous 32 GB DRAM 32 GB PCM PCM parameters 4X read latency 4X write energy 10 M writes/cell Hybrid 1 GB DRAM 32 GB PCM
23
SLIDE 24
24
PCM lifetimes
10 20 30 40 Lifetime in years PCM-Only KG-N KG-W
9 17 PCM alone is not practical PCM lasts more than 10 years with KG-W
SLIDE 25
25
EDP reduction compared to DRAM
- 80
- 40
40 80 % reduction in EDP PCM-Only KG-N KG-W
EDP : Energy Delay Product KG-W has 35% better EDP than DRAM-Only Higher is better 4 cores
SLIDE 26
26
Emulation on NUMA hardware
D R A M D R A M
DRAM: Socket 0 CPU CPU PCM: Socket 1
Modify JVM to divide heap in DRAM and PCM Use Intel perf monitor to measure writes
D R A M D R A M
SLIDE 27
27
PCM write rates on NUMA hardware
KG-N reduces write rate by 3.8X over PCM-Only KG-W reduces write rate by 1.9X over KG-N
0.0 0.5 1.0 1.5 DaCapo Pjbb GraphChi Avg Write rate in GB/s PCM-Only KG-N KG-W
130 MB/s
SLIDE 28
28
Crystal Gazer: Profile-Driven Write-Rationing Garbage Collection for Hybrid Memories
SLIDE 29
Promising to monitor heaps at a fine granularity
29