Transcript Cloud ISTC
Efficient Mixed-Platform Clouds Phillip B. Gibbons, Intel Labs Michael Kaminsky, Michael Kozuch, Padmanabhan Pillai (Intel Labs) Gregory Ganger, David Andersen, Garth Gibson (Carnegie Mellon) NSF Workshop on Sustainable Energy Efficient Data Management May 2, 2011 1 Cloud Computing & Homogeneity • In near future, significant fraction of all data analysis and data storage will occur in the cloud • Traditional data center goal: Homogeneity + Reduce administration costs: maintenance, diagnosis, repair + Ease of load balancing CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU CPU Disk Mem Disk Mem Disk Mem Disk Mem … CPU CPU CPU CPU Disk Mem Ideal: single Server Architecture tailored to the workload 2 Homogeneity: Challenges • No single workload: Mix of customer workloads – Computation-heavy apps (powerful CPUs, little I/O BW) – Random I/O apps (I/O latency bound) – Streaming apps (I/O BW bound, little memory) – Memory-bound apps – Apps exploiting hardware assists such as GPUs • Common denominator Server Architecture falls short – E.g., Two orders of magnitude loss in energy efficiency (see example on next slide) 3 FAWN: Fast Array of Wimpy Nodes • For key-value stores, FAWN provides 120X more queries per Joule than traditional server • FAWN great for some workloads, terrible for others Homogeneity 4 New Goal: Specialization •Specialization is fundamental to efficiency – No single platform best for all application types – e.g., huge efficiency gains in FAWN – Called division of labor in sociology (see also, bees) •Cloud computing must embrace specialization – and consequent heterogeneity and change-over-time Specialization is fundamental to sustainable energy-efficient data management 5 Efficient Mixed-Platform Clouds Cloud in 2020 will need… • Infrastructure purposely composed of many platform types, some general-purpose and some specialized to particularly important application types • Infrastructure embraces heterogeneity by design • Nimble incorporation of new technologies is enabled by explicitly aiming for heterogeneity – E.g., solid state RAM and accelerators 6 Efficient Mixed-Platform Cloud Research Agenda • Develop specializations motivated by important application types • Algorithms/frameworks for exploiting specializations • Making applications able to work on varied platforms – And automatically mapping them to best platform, accounting for where the data is • Explore disruptive impact of new technologies – integration into systems, exploitation by applications • Data management in mixed-platform cloud Our progress to date on specializations: See FAWN [SOSP’09], Hi-Spade [Sigmod’10,Sigmod’11], PCM-DB [CIDR’11] projects 7 Coming Soon: Intel Science and Technology Center on Cloud Computing (ISTC-CC) • Pending approvals, legal agreements, etc • $2.5M / year for 3-5 years • Homed at Carnegie Mellon • 4 Intel researchers Research Agenda 8 Back Up Slides 9 Defining Cloud Computing… • Easy to get mired in defining cloud computing – we really want to avoid doing so (again ) • NIST ended up with a 2-page definition – here’s their 15th version, for reference: • Is it … – Amazon Web Services (EC2, S3, etc.) ? – Google Apps + Chrome ? – Private clouds based on VMware/Eucalyptus/etc ? – Hadoop / MapReduce ? – NoSQL DBs (Cassandra, etc.) ? • All are examples of broad collection of trends 10 Cloud in 2020? •Huge range of uses, exploiting … – shared, managed resources – needs to be massive scale, efficient, automated, trustworthy – availability of interesting data – needs to support BIG DATA, sensor data, mining of both – convenient on-demand access from anywhere – needs to be elastic, easy-to-use, location-independent 11