nasa nebula · 2010. 5. 7. · •evented, parallel web service interface •multicast file...

NEBULA Cloud Computing Platform 2010

Cloud Storage for Open Science

NASA Nebula

NEBULA Cloud Computing Platform 2010

I am not a scientist.

NEBULA Cloud Computing Platform 2010

NEBULA Cloud Computing Platform 2010

This is not your mother’s storage array.

NEBULA Cloud Computing Platform 2010

Challenges

• Be all things to all people

• Be cheaper

• Be faster

• …and still be reliable (enough)

NEBULA Cloud Computing Platform 2010

Free Market Closed Market

NEBULA Cloud Computing Platform 2010

Approach

• Copy the private sector

• Ignore vendors

• Work from first principles

• Test, measure, and repeat

• Fail fast

NEBULA Cloud Computing Platform 2010

NEBULA Cloud Computing Platform 2010

Find your cost leverage…

NEBULA Cloud Computing Platform 2010

..and match it.

NEBULA Cloud Computing Platform 2010

POSIX / POSIX-- / KVS

• Fast, temporary scratch disks

• “Pretty” big, “pretty” reliable NAS

• Very, very large KVS (by file size AND file count)

NEBULA Cloud Computing Platform 2010

Our conclusions

• Containerized Data Centers

• Open standards: HTTP, WebDAV, Ethernet

• Monitoring-in-depth

• Do battle with every SPoF

• Recompute

NEBULA Cloud Computing Platform 2010

When you have a hammer…• Solve it with software

• Solve it with hardware

• Solve it with $$$

NEBULA Cloud Computing Platform 2010

Custom solutions (ongoing)

• Evented, parallel web service interface

• Multicast file replication (IPv4 and IPv6)

• MapReduce-supporting Object Store

• Alternate filesystems

• FusionIO

NEBULA Cloud Computing Platform 2010

The Gauntlet

NEBULA Cloud Computing Platform 2010

Your DATA

• Raw Data

• Metadata

Storage Interface

• POSIX

• Web Services

NAS Protocol

• AoE (L2)

• iSCSI / NFS (L3)

File System

• Distributed (Stripe Size)

• Local FS (Block Size)

• Partition Table (maybe)

RAID Volume

• Block size

• Drive Count

RAID Controller

• BBU

• Cache Size

Raw Disk

• Interface (SAS / SATA / SSD)

• Reliability

NEBULA Cloud Computing Platform 2010

Stop passing the buck

• Stated risk: 10^16

• Actual risk: 1 in 72

NEBULA Cloud Computing Platform 2010

References

• http://lauraparkin.wordpress.com/2009/07/11/fail-early-fail-fast/

• http://nebula.nasa.gov