Doctoral Thesis: Application Workload Prediction and Placement in Cloud Computing Systems

SHARE:

Event Speaker: 

Katrina LaCurts

Event Location: 

32-G449 (Kiva)

Event Date/Time: 

Friday, April 11, 2014 - 2:00pm

 

Cloud computing has become popular in recent years.  Companies such as

Amazon and Microsoft host large datacenters of networked machines

available for users to rent.  These users are varied: from individual

researchers to large companies.  Their workloads range from short,

resource-intensive jobs to long-running user-facing services.

 

As cloud environments become more heavily used, provisioning the

underlying network becomes more important.  Previous approaches to

deal with this problem involve changing the network infrastructure,

for instance by imposing a particular topology or creating a new

routing protocol.  While these techniques are interesting and

successful in their own right, we ask a different question: How can we

improve cloud networks without changing the network itself?  This

question is motivated by two desires: first, that customers be able to

improve their application's performance without necessarily involving

provider, and second, that our techniques be immediately applicable to

today's cloud networks.

 

This dissertation presents an end-to-end system, Cicada, which

improves application performance on cloud networks.  Cicada tackles

two problems: how to model and predict an application's workload, and

how to place applications on machines in the network.

 

Cicada can be used by either cloud providers or customers.  When used

by a cloud provider, Cicada enables the provider to offer certain

network performance guarantees to its customers.  These guarantees

give customers the confidence to use cloud resources when building

their own user-facing applications (as there is no longer a risk of

the cloud under-provisioning for the customer's network needs), and

allow providers to better utilize their network.  When used by

customers, Cicada enables customers to satisfy their own objectives,

such as minimizing the completion time of their application.  To do

this, Cicada exploits variation in the underlying cloud network to use

the fastest paths most frequently.  This requires an extension to

Cicada, called Choreo, which performs quick, accurate, client-side

measurement.

 

We evaluate Cicada using data we collected from HP Cloud, a deployed

network with real users.  Cicada's workload prediction algorithm

outperforms the existing state-of-the-art by 90%.  Its placement

method reduces application completion time by an average of 22%--43%

(maximum improvement: 79%) when applications arrive in real-time, and

doubles inter-rack utilization in certain datacenter topologies.

These results and others herein corroborate our thesis that

application performance can be improved in cloud networks without

making any changes to the network itself.

 

 

Thesis Supervisor: Hari Balakrishnan