By Jeroen Klaver & Roel van der Jagt Background Research - PowerPoint PPT Presentation

By Jeroen Klaver & Roel van der Jagt

 Background  Research question  Test approach  SURFnet design  Conclusion  Future work 2

 SURFnet is the Dutch NREN  Storage explosion 3

 Add more storage 4

 And more...  But adding includes high costs and new bottlenecks 5

 SURFnet thinks there is a smarter way 6

7

8

9

 What infrastructure and open source tools provide SURFnet or the participants a scalable and distributed any-kind-storage solution? 10

 Coda  Lustre  GlusterFS  XtreemFS  Ceph  PVFS  MooseFS  Hadoop 11

 First stage  Null test  Second stage  Architectures in the lab environment  Third stage  Architectures in practice 12

13

Iozone - Lustre file system benchmark 90000 80000 Speed (KBytes/sec) 70000 60000 50000 40000 30000 Lustre avg 20000 Null test avg 10000 0 256 8192 256 8192 256 8192 256 8192 Sequential Sequential Random Random read write read write Block size (KBytes) 14

Iozone - Lustre with latency 90000 80000 Speed (KBytes/sec) 70000 60000 50000 40000 30000 No added latency avg 20000 Latency (20ms RTT) avg 10000 0 256 8192 256 8192 256 8192 256 8192 Sequential Sequential Random Random read write read write Block size (KBytes) 15

16

Iozone - Coda file system benchmark 70000 60000 Speed (KBytes/sec) 50000 40000 30000 20000 Coda avg Null test avg 10000 0 256 8192 256 8192 256 8192 256 8192 Sequential Sequential Random read Random write read write Block size (KBytes) 17

Iozone - Coda with latency 60000 50000 Speed (KBytes/sec) 40000 30000 20000 No added latency avg Latency (20ms RTT) avg 10000 0 256 8192 256 8192 256 8192 256 8192 Sequential Sequential Random Random read write read write Block size (KBytes) 18

Iozone - File system benchmark 90000 80000 Speed (KBytes/sec) 70000 60000 50000 40000 Lustre avg 30000 Coda avg 20000 Null test avg 10000 0 256 8192 256 8192 256 8192 256 8192 Sequential read Sequential Random read Random write write 19 Block size (KBytes)

Bonnie++ - Seeks 700 600 500 Seeks (seeks/sec) 400 Lustre Coda 300 Null test 200 100 0 min avg max 20

Iozone - Coda vs Lustre with latency 90000 80000 Speed (KBytes/sec) 70000 60000 Lustre No added latency avg 50000 40000 Lustre Latency (20ms RTT) 30000 avg 20000 Coda No added latency avg 10000 0 Coda Latency (20ms RTT) 256 8192 256 8192 256 8192 256 8192 avg Sequential Sequential Random Random read write read write Block size (KBytes) 21

 Requirements  Scalable  Available  Durable  Performance  Dynamic  Cost effective  Generic interfaces  Open protocols  Geographic dispersion  Deployment \ Maintenance  Security features 22

 Requirements  Scalable  Available  Durable  Performance  Dynamic  Cost effective  Generic interfaces  Open protocols  Geographic dispersion  Deployment \ Maintenance  Security features 23

 Coda:  Hard to configure  Can not handle large files  Promising architecture  Low latency impact  Lustre:  DRBD is needed for replication / latency reduction  Latency has impact  No multi-tier  Promising future 24

 Review of GlusterFS  Test environment without VMWare ESXi  Further tuning and configuring of Coda  Coda and ZFS tests  Lustre and DRBD tests 25

(Source: mastergoogle.com) 26

27

28

29

30

31

32

33

 Coda min avg max 27,6 28 28,4 No added latency Latency (20ms RTT) 9,5 9,8 10  Lustre min avg max 151 153 155 No added latency Latency (20ms RTT) 28,7 29 29,2 34