Machine Placement and Migration in Cloud Datacenters Authors: - - PowerPoint PPT Presentation

Cache Contention Aware Virtual Machine Placement and Migration in Cloud Datacenters

Authors: Liuhua Chen, Haiying Shen and Stephen Platt Presenter: Haiying Shen

IEEE ICNP

November 8-11, 2016 Singapore

2

An effective VM allocation algorithm should allocate as many VMs as possible to a PM i) meeting explicit resource requirements (CPU, memory) ii) minimizing contentions on Last Level cache Many previous VM allocation or migration methods provide a metric to choose destination PM and migration VM to handle objective i) but neglect objective ii).

Objective

3

Objective

VM1 PM VM2 Last Level Cache

Performance degradation due to shared cache

• Reduce cache interference

in VM consolidation

VM2 VM1

4

• A brief review of cache hierarchy
• VM cache performance degradation prediction
• VM placement and migration algorithm
• Experimental results
• Conclusion with future directions

Overview

5

A brief review of cache hierarchy

A B C D [A,B,A,B,C,D,D,B]

LRU stack C1 C2 C3 C4 C1 increases 1 C3 increases 1

64KB L1 256KB L2

Core 1

8MB L3 Cache

64KB L1 256KB L2

Core 1

64KB L1 256KB L2

Core 1

64KB L1 256KB L2

Core 1

stack distance profile

6

Stack distance profile

The number of accesses

Cache Hits Cache Misses C1 CA C>A … Stack distance counters fi = {C1,C2,...,CA,C>A}, Cd counts the number of hits to the line in the dth LRU stack position and C>A counts the number of cache misses

7

Stack distance profile (cont.)

Number of accesses Cache Hits Original Cache Misses C1 CA C>A … Stack distance counters Extra Cache Misses

8

When VM i and VM j compete for a cache line (which is the dth position in the LRU stack), the probability of VM i “winning” the competition is proportional to the number of accesses to this cache line of VM i, but reversely proportional to the total number of accesses to this cache line of the two VMs.

Cache Contention Prediction

Extend to multiple VMs

9

The new stack distance profile of VM i can be estimated by

Cache Contention Prediction (cont.)

10

Sensitivity is a measure of how much a VM will suffer when cache space is taken away from it due to contention. Intensity is a measure of how much a VM will hurt others by taking away their space in a shared cache. The degradation of co-scheduling vi and vj together is the sum of the performance degradation

• f the two VMs

Performance Degradation Prediction

11

Objective: minimize the total pain of the co-location of the new VMs with the existing VMs

VM placement and migration algorithm

Define it as an

• ptimization problem

Transform it to an integer linear programming Use lpsolve 5.5 tool to find optimal solution

12

The computational complexity of the above method is very high, especially for a relatively large number of VMs. We propose a heuristic VM placement and migration algorithm.

• VM placement: allocates each VM to a PM that leads to the minimum total performance

degradation.

• VM migration: select a VM which generates the maximum pain with other co-located

VMs in the PM to migrate out.

VM placement and migration algorithm (cont.)

13

Simulation:

• CloudSim (extended to model LLC

contention)

• Use trace to determine profiles
• 1000+ PMs
• 4000 VMs

Experimental results

Real testbed:

• High-performance computing (HPC)

cluster

• Each VM run NPB suite workload
• - NAS Parallel Benchmark (NPB) suite
• 20 PMs
• 120 VMs

Comparison algorithms: cache unaware (Random), classification based (Animal), miss rate based (MissRate) Our algorithm: CacheVM

14

Model validation

20 40 60 80 100 5 10 15

Prediction error (%) CDF of prediction error (%)

This result confirms that the proposed model achieves a high accuracy in predicting cache behaviors. (Cache misses predicted by the model - Cache misses collected by the simulator)/Cache misses collected by the simulator)

15

Comparison with the Optimal Algorithm

0.0E+00 1.0E+05 2.0E+05 3.0E+05 4.0E+05 20 30 40

Total # of misses The number of VMs

Random Animal MissRate CacheVM Optimal 1.0E+06 1.0E+08 1.0E+10 1.0E+12 20 30 40

Total time (ns) The number of VMs

Random Animal Missrate CacheVM Optimal

Optimal<CacheVM<Animal<Random MissRate increases faster MissRate<CacheVM<Random<Animal<<Optimal 20 PMs 20 PMs

16

Simulated performance with real trace

0.0E+00 5.0E+06 1.0E+07 1.5E+07 2000 3000 4000

Total # of misses The number of VMs

Random Animal MissRate CacheVM 0.0E+00 2.0E+06 4.0E+06 6.0E+06 8.0E+06 1.0E+07 1.2E+07 1.4E+07 Scale 1 Scale 2 Scale 3

Total # of misses Different scales

Random Animal MissRate CacheVM

CacheVM<Animal≈Random<MissRate CacheVM<Animal<Random<MissRate

(1000 VMs, 750 PMs), (2000 VMs, 1500 PMs), (4000 VMs, 3000 PMs)

2000 PMs

17

Performance on real testbed

0.0E+00 5.0E+03 1.0E+04 1.5E+04 2.0E+04 20 60 90 120

The number of VMs

Random Animal MissRate CacheVM

Total execution time (s)

0.0E+00 1.0E+04 2.0E+04 3.0E+04 4.0E+04 5.0E+04 6.0E+04 20 60 90 120

The number of VMs

Random Animal MissRate CacheVM

Total throughput (Mop/s)

CacheVM<MissRate<Animal<Random Random<Animal<MissRate<CacheVM Varied VMs from 20 to 120 and allocated them to 20 PMs

18

Performance on real testbed (cont.)

0.0 0.5 1.0 1.5 2.0 20 60 90 120

Normalized time

The number of VMs Random Animal MissRate CacheVM 0.0 0.3 0.6 0.9 1.2 1.5 20 60 90 120 The number of VMs Random Animal MissRate CacheVM

Normalized throughput

CacheVM<Animal<MissRate<Random Random<Animal<MissRate<CacheVM

19

• Proposed a cache contention aware VM performance degradation prediction algorithm.
• Formulated a cache contention aware VM placement problem.
• Transformed this problem to an integer linear programming (ILP) model and solved it.
• Proposed a heuristic cache contention aware VM placement and migration algorithm
• Conducted trace-driven simulation and real-testbed experiments to evaluate CacheVM.

Future work: develop a decentralized version of the proposed algorithm.

Conclusion and future work

20

Thank you! Questions & Comments?

Haiying Shen hs6ms@v @virgin ginia.e .edu du Pervas asive ive Communi unica catio tion n Laborato tory ry Univers rsity ty of Virginia