Pelican: A building block for exascale cold data storage Shobana Balarishnan, Richard Black, Austin Donnelly, Paul England, Adam Glass, Dave Harper, Sergey Legtchenko, Aaron Ogus, Eric Peterson, Antony Rowstron Microsoft Research, Microsoft Adapted from their OSDI and SDC slides. EECS 582 – W16
Background: cold data in the cloud EECS 582 – W16
Background: cold data in the cloud EECS 582 – W16
Right-provisioning EECS 582 – W16
Pelican: rack-scale appliance for cold data EECS 582 – W16
Prototype EECS 582 – W16
Impact of right provisioning on resources x12 x6 EECS 582 – W16 x16
Data placement: maximizing request concurrency EECS 582 – W16
Data placement: maximizing request concurrency EECS 582 – W16
Data placement: maximizing request concurrency EECS 582 – W16
IO scheduling: “spin up is the new seek” EECS 582 – W16
Challenges of right-provisioning EECS 582 – W16
Evaluation FP: full provisioning EECS 582 – W16
Throughput EECS 582 – W16
Time to first byte EECS 582 – W16
Power consumption EECS 582 – W16
Conclusion Pros and Cons Reduce capital cost and operating cost Meet requirements from cold data workload Sensitive to hardware changes Manually handle many constraints EECS 582 – W16
x4 x2 Related Works Amazon Glacier Facebook cold storage datacenter Two billion photos Maybe right-provisioning idea Blue-ray optical disk x4 x2 EECS 582 – W16
Q&A Thank You! EECS 582 – W16