Reciprocal Resource Fairness: Towards Cooperative Multiple-Resource Fair Sharing in IaaS Clouds

School of Computer Engineering

Nanyang Technological University, Singapore

Haikun Liu and Bingsheng He

Current IaaS Cloud Model (T-shirt)

VM instance

CPU (EC2 comp. unit)

Memory (GB)

Storage (GB)

Price ($/hour)

Small 1 1.7 160 0.06

Medium 2 3.75 410 0.12

Large 4 7.5 850 0.24

Ext Large 8 15 1690 0.48

• Popular Cloud providers sell VM instances with fixed capacity (T-shirt).

• Charge users based on resources used over time (Pay-as-you-use).

• Horizontal resource scaling (Scale-out).

Disadvantages of T-shirt Model

• Tenants’ resource demands are

heterogeneous (NSDI’11). – Tenants have different resource demands. – A tenant’s demand is changing over time.

• Resource utilization is a critical problem in such pay-as-you-use environments. – Cloud providers waste resource.

higher operating cost and less revenue.– Cloud tenants waste their money.

Static resource allocation (T-shirt Model) causes resource underutilization or bad application performance.

To Share or Not To Share?

• Resource Utilization = $• Resource Sharing can improve resource efficiency.

– Allow underloaded tenants to release resources to other tenants. – Allow overloaded tenants to temporarily use more resources

(from others).

• Virtualization technologies already provide enough technical supports for resource sharing.– CPU, I/O multiplexing (time-sharing)– Memory Overcommit (ballooning, hotplugging)

A New Resource Alloc. Model

• Time-sharing Model– Compatible with current cloud

interface (static billing)– Allow dynamic resource scaling for

VMs in a fine-grained manner

• Challenges: fairness – Free-riding– Lying– Economic fairness

Tenants

Resource pool

If the fairness problem is not solved, tenants should not have incentive to share resource.

Economical Fairness: Resource-as-you-pay • The total value of resources the tenant received should

be proportional to her payment. • This is a Service-Level Agreement (SLA).

$ 50

$ 50$ 50

A:

B:

MemCPU

A: 50% B: 50%B: 50%

MemCPU

Value of ResourcePayment

Existing Fair Policies

• State-of-the-art: – Weighted Max-Min Fairness (WMMF): always select the user

with the minimum demand/share ratio every time.– Dominant Resource Fairness (DRF): always maximize the

smallest dominant share of users in a system (NSDI’11).

• Disadvantages of resource allocation for multiple resource types:– Free-riding– Lying– Economical fairness

Problems of T-shirt, WMMF, DRF

VMs VM1 VM2 VM3 Total

Initial Shares <500, 500> <500, 500> <1000, 1000> <2000, 2000>

Demands <6 GHz, 3 GB> <8 GHz, 1 GB> <8 GHz, 8 GB> <22 GHz, 12GB>

T-shirt Allocation

<5 GHz, 2.5GB> <5 GHz, 2.5GB> <10 GHz, 5 GB> actually used <18 GHz, 8.5GB>

WMMF Allocation

<6 GHz, 3 GB> <6 GHz, 1 GB> <8 GHz, 6 GB> <20 GHz, 10 GB>

WDRF dominant share

6/20 = 3/10 8/20 CPU 8/(10*2) RAM 100%

WDRF Allocation

<6 GHz, 3 GB>

<7 GHz, 1 GB> <7 GHz, 6 GB> <20 GHz, 10 GB>

• Example: Three VMs share total 20 GHz CPU and 10 GB RAM.

Unused <0 GHz, 0 GB> <0GHz, 1.5GB> <2 GHz, 0 GB>

• Challenges: can we find a fair sharing policy that satisfies the following properties? – Sharing Incentive– Gain-as-you-contribute Fairness – Strategy-proofness

• Solution: Reciprocal Resource Fairness– The basic idea is to allow flexible resource allocation

to VMs while keeping the total resource value unchanged.

Our Work: RRF

Reciprocal Resource Fairness (RRF)• Hierarchical and complementary mechanisms：

– Inter-tenant Resource Trading (IRT)– Intra-tenant Weight Adjustment (IWA)

PM

VM1

TenantA

TenantM

RT

VMnWA VM1

VMn

WA

…

Resource Alloc. Model

• Normalize different types of resources based on their market price.– A tenant’s asset is the aggregate shares of all resource types.

• Resource allocation model:– Payment Shares Resources– A VM’s resource share reflects its priority relative to other VMs.– Resource allocation is determined by shares, tenant’s payment.

Inter-tenant Resource Trading

• Tenant’s gain from other tenants should be proportional to her contribution.

MemCPUMem

CPU200

MemCPU

100

MemCPU

300

contributions

A B C D

Inter-tenant Resource Trading• Tenant’s gain from other tenants should be proportional

to her contribution.

MemCPUMem

CPU

200

MemCPU

100

MemCPU

200

100

Contribution of Memory : A:B= 200:100

Gain of CPU: A:B= 200: 100

A B C D

Discussion

• Comparison of WMMF, DRF, RRF

Proof sketches are in the paper.

Evaluation• Testbed:

– implemented RRF on Xen 4 and deploy the prototype in a cluster with 10 nodes.

• Benchmark: – TPC-C, RUBBoS, Kernel-build, Hadoop

• Workloads:– Stable, cyclical on-off, heavy

and light – Application are running in one or

more VMs. • Methodology:

– T-shirt, WMMF, DRF

The ratio of total resource demandto total initial share

Economic Fairness

• RRF can guarantee 95% economic fairness for multi-resource sharing among multi-tenants.

Application Performance

• RRF delivers 45% app performance improvement to tenants compared to T-shirt model.

VM density vs. App Performance

• RRF improves VM density than T-shirt model by 2.2X at the

expense of around 15% performance penalty.

Performance Overhead

• RRF causes reasonable CPU load on the host machine;• RRF causes negligible performance overhead on guest VMs.

Conclusion

• A new resource sharing model for IaaS clouds.

• Reciprocal resource fairness with two complementary mechanisms: inter-tenant resource trading and intra-tenant weight adjustment.

• RRF can guarantee economical fairness, and Improve resource efficiency and application performance.