SLIDE 1
NUMA-Friendly Stack (using Delegation and Elimination)
Irina Calciu Justin Gottschlich Maurice Herlihy
HotPar ‘13
1
Trends for Future Architectures
2
NUMA-Friendly Stack (using Delegation and Elimination) Irina Calciu - - PowerPoint PPT Presentation
NUMA-Friendly Stack (using Delegation and Elimination) Irina Calciu - - PowerPoint PPT Presentation
NUMA-Friendly Stack (using Delegation and Elimination) Irina Calciu Justin Gottschlich Maurice Herlihy HotPar 13 1 Trends for Future Architectures 2 Uniform Memory Access (UMA) 3 Non-Uniform Memory Access (NUMA) NUMA NODE (multiple
SLIDE 2
SLIDE 3
Uniform Memory Access (UMA)
3
SLIDE 4
Non-Uniform Memory Access (NUMA)
(interconnect) NUMA NODE (multiple cores, shared Last Level Cache) NUMA NODE (multiple cores, shared Last Level Cache) NUMA NODE (multiple cores, shared Last Level Cache) NUMA NODE (multiple cores, shared Last Level Cache) Cache coherency maintained between caches on different NUMA nodes
4
SLIDE 5
Overview
- Motivation
- Algorithms
- Results
- Conclusions
5
SLIDE 6
Delegation
NUMA node 0 NUMA node 1 Clients Clients SEQ STACK Server
6
SLIDE 7
Delegation
NUMA node 0 NUMA node 1 Server Client 5 Client 6 Client 7 Client 8 Slots Client 1 Client 2 Client 3 Client 4 Slots Loops through all slots SEQ STACK
7
SLIDE 8
Elimination, Rendezvous
8
SLIDE 9
Local Rendezvous
NUMA node 0 NUMA node 1 STACK
9
SLIDE 10
Delegation + Elimination
NUMA node 0 NUMA node 1 Clients Clients SEQ STACK Server
10
SLIDE 11
Delegation + LOCAL Elimination
NUMA node 0 NUMA node 1 Clients Clients SEQ STACK Server
11
SLIDE 12
Effect of Elimination
Throughput (Better) 50% push 50% pop 90% push 10% pop
12
SLIDE 13
Effect of Delegation
Throughput (Better) 50% push 50% pop 90% push 10% pop
13
SLIDE 14
Number of Slots
Throughput (Better) 50% push 50% pop 90% push 10% pop
14
SLIDE 15
Workloads: Balanced vs. Unbalanced
Throughput (Better) 50% push 50% pop 70% push 30% pop
15
SLIDE 16
Advantages
- Memory and cache locality
- Reduced bus traffic
- Increased parallelism through elimination
16
SLIDE 17
Drawbacks
- Communication cost between clients and
server thread
- Insignificant compared to the benefits
- Serializing otherwise parallel data structures
- Parallelism through elimination
- Elimination opportunities decrease as
workload more unbalanced
17
SLIDE 18
Open Questions
- Are there other data structures where we can use
delegation and elimination?
- Are there data structures where direct access is
much better?
- What can we do for those data structures?
18
SLIDE 19
Thank you! Questions?
19
SLIDE 20
References
- A Scalable Lock-free Stack Algorithm
http://www.inf.ufsc.br/~dovicchi/pos-ed/pos/artigos/p206- hendler.pdf
- Flat Combining and the Synchronization-Parallelism Tradeoff
http://www.cs.bgu.ac.il/~hendlerd/papers/flat-combining.pdf
- Fast and Scalable Rendezvousing
http://www.cs.tau.ac.il/~afek/rendezvous.pdf
20
SLIDE 21
Cache to Cache Traffic
Better
21
SLIDE 22
Coefficient of Variation
Better
22
SLIDE 23
Flat Combining
23
SLIDE 24
Delegation
CLIENT Find corresponding slot (by NUMA node and cpuid) Post message Wait for response Get response SERVER Loop through all slots: If slot has message: Take message Process message Send response Time
24
SLIDE 25
Delegation
CLIENT Find corresponding slot (by NUMA node and cpuid) try_elimination: if (eliminate) return Post message Wait for response Get response else try_elimination SERVER Loop through all slots: If slot has message: Take message Process message Send response Time
25
SLIDE 26
Delegation
CLIENT Find corresponding slot (by NUMA node and cpuid) try_elimination: if (eliminate) return if (Acquire slot lock) Post message Wait for response Get response Release slot lock else try_elimination SERVER Loop through all slots: If slot has message: Take message Process message Send response Time
26
SLIDE 27
Open Questions
- Performance
- Scalability
- Power
27