Parallel Computing in the Cloud

Slide 1

Slide 1 text

Parallel Computing in the Cloud Dave Fellows (CTO) Ivan Towlson (Chief Architect)http://www.greenbutton.com/ [email protected] [email protected]

Slide 2

Slide 2 text

No content

Slide 3

Slide 3 text

www.greenbutton.com Parallel Architectures • Shared memory: lots of cores in one machine • HPC cluster: lots of cores with fast communication between them • “Army of Davids” – Server farm or data centre populated with commodity machines – Crowdsourcing (Folding@Home etc.)

Slide 4

Slide 4 text

www.greenbutton.com Parallel Cloud • Recruiting an army, even of Davids, is expensive and time-consuming • Then you have to provide them with pay and rations even if they’re just sitting around • Cloud benefits: – Zero lead time – Zero capital requirement – Zero costs when not in use

Slide 5

Slide 5 text

www.greenbutton.com Parallel Cloud at GreenButton • Architecture • Tools • Experiences and Lessons Learned

Slide 6

Slide 6 text

www.greenbutton.com

Slide 7

Slide 7 text

www.greenbutton.com Web Service Job submission Controller Job allocation Instance provisioning Status monitoring Processor Runs compute tasks Billing Monitoring Notification Authorisation GreenButton.com Mission Control Web app Web services (user/account) Asset streaming Apps

Slide 8

Slide 8 text

www.greenbutton.com Web Service Controller Processor Tasks placed on queue Job record written to cloud storage Requests next job Monitors job list Starts up and shuts down processor instances according to demand Tasks pulled off queue Task processor Job splitter

Slide 9

Slide 9 text

www.greenbutton.com Parallel Cloud at GreenButton • Architecture • Tools • Experiences and Lessons Learned

Slide 10

Slide 10 text

www.greenbutton.com

Slide 11

Slide 11 text

www.greenbutton.com

Slide 12

Slide 12 text

www.greenbutton.com $1360/hr

Slide 13

Slide 13 text

www.greenbutton.com Parallel Cloud at GreenButton • Architecture • Tools • Experiences and Lessons Learned

Slide 14

Slide 14 text

www.greenbutton.com Animation and Rendering • Bandwidth challenges – Data can be very large for TV/Movie quality • Environmental complexities • Licensing and dynamic provisioning – Animation software is high-value and proprietary • Security (sometimes!)

Slide 15

Slide 15 text

www.greenbutton.com Scientific Computing • MPI challenges – Inter-node bandwidth and latency – Managing dynamic nodes as a virtual cluster instead of independent actors – Managing failure • check pointing • Live migration with virtualization – Virtualization overhead?

Slide 16

Slide 16 text

www.greenbutton.com Financial Computing • There’s an awful lot of Excel in finance! – Deployment challenge – Partitioning challenges • Data sensitivity

Slide 17

Slide 17 text

www.greenbutton.com Oil & Gas

Slide 18

Slide 18 text

www.greenbutton.com

Slide 19

Slide 19 text

www.greenbutton.com Oil & Gas • Bandwidth, bandwidth and more bandwidth – 50 terabyte files defeat online upload – Scalable file system (Gluster) • Seismic imaging – visualization – Entire workflow moving to cloud

Slide 20

Slide 20 text

API / Controller VIZ Glacier

Slide 21

Slide 21 text

www.greenbutton.com Parallel Cloud Futures • High memory instances – We have lots of cores – RAM is harder to come by • Infiniband / RDMA – Virtualisation challenges • GPUs • “Bare metal cloud”

Slide 22

Slide 22 text

www.greenbutton.com Conclusion • Challenges – Bandwidth in and out of cloud – Latency between cloud nodes – Limits on node sizing – Licensing – Immediacy – spin-up time – Cloud lock-in – Governance – operational budget control

Slide 23

Slide 23 text

www.greenbutton.com Conclusion • Huge benefits for many parallel workloads • Choose candidate applications wisely • Don’t roll your own cloud solution http://www.greenbutton.com/ [email protected] [email protected]