cloud- tm: a distributed transactional memory platform for the cloud
DESCRIPTION
Cloud- TM: A distributed transactional memory platform for the Cloud. Paolo Romano INESC ID Lisbon, Portugal. At a glance. Cloud computing: the vision. lower barriers to entry and capital costs via usage-based pricing schemes - PowerPoint PPT PresentationTRANSCRIPT
technologyfrom seed
Cloud-TM:A distributed transactional memory platform for the CloudPaolo RomanoINESC ID Lisbon, Portugal
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
At a glance
21st Plenary EuroTM Meeting, Paris 19-20 May 2011
Cloud computing: the vision
• lower barriers to entry and capital costs via usage-based pricing schemes
• minimize operating costs & carbon footprint via elastic resource provisioning
• achieve unprecedented scalability levels
31st Plenary EuroTM Meeting, Paris 19-20 May 2011
Project Motivations
• Cloud computing is at the peak of its hype…
• How to materialize the vision and maximize actual productivity?
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
SIMPLIFYING THE DEVELOPMENT AND ADMINISTRATION OF CLOUD APPLICATIONS
Key Goals
Develop an open-source middleware platform for the Cloud:
1. Providing a simple and intuitive programming model:➭ hide complexity of distribution, persistence, fault-tolerance➭ let programmers focus on differentiating business value
2. Minimizing administration and monitoring costs:➭ automate elastic resource provisioning based on applications
QoS requirements
3. Minimize operational costs via self-tuning➭ maximizing efficiency adapting consistency mechanisms upon
changes of workload and allocated resources
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
The Cloud-TM Programming Paradigm:Background
• Transactional Memories (TM):– replace locks with atomic transactions in multi/many core systems– hide away synchronization issues from the programme– drastically simplify development of parallel applications
• Distributed Transactional Memories (DTM):– extends the TM abstraction over the boundaries of a single machine– avoid performance pitfalls of Distributed Shared Memory via
speculation:• no need to synchronize every and each read/write operation• batch consistency actions at commit time• have been shown to scale up to hundreds of node (Sinfonia,
ClusterSTM)
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
The Cloud-TM Programming Paradigm:Elastic Distributed Transactional Memory
• Elastic scale-up and scale-down of the DTM platform:– data distribution policies minimizing reconfiguration overhead– auto-scaling based on user defined QoS & cost constraints
• Transparent support for fault-tolerance via data replication:– self-tuning of consistency protocols driven by workload changes
• Language level support for:– transparent support of object-oriented domain model (incl. search)– highly scalable abstractions– parallel transaction nesting in distributed environments
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
Architectural Overview
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
Autonomic adaptation at play
1st Plenary EuroTM Meeting, Paris 19-20 May 2011
Conclusions
• Whole new breed of research challenges:– elasticity:
• self-tuning as an essential requirement
– non-uniform transaction synchronization costs:• multi-core rack data-center cloud federation
– unprecedented scalability challenge
101st Plenary EuroTM Meeting, Paris 19-20 May 2011
MULTICORES are to STMs... as CLOUD are to DSTMs
Project’s timeline
Internet of Services Collaboration Meeting – Bruxelles 19-20 October 2010
June 2011
Dec2010 June
2013
Mar2013
Dec2012