Steve Marks.
PASIG — Preservation and Archiving Special Interest Group 2015 Meeting.
https://libraries.ucsd.edu/chronopolis/pasig/agenda_2/index_agenda.html
1. The Ontario Library Research
Cloud
Steve Marks
Scholars Portal / Ontario Council of University Libraries
(University of Toronto)
2. What is Scholars Portal
Shared technology service of the 21 university libraries of
the Ontario Council of University Libraries founded in
2003
Provides content aggregation and preservation services for
member libraries
Journals – 16,000 and 38 M articles
Books – 610,000 ebooks
GeoPortal – GIS Data
ODESI – Numeric Data
Dataverse – Research Data
9. MTCU PIF Proposal
Nine partner libraries from OCUL; three year project
University of Toronto as financial lead
Develop a 1.2PB object storage service for partners
Provide subscription storage services to other OCUL
libraries
Develop interfaces with library repository applications
Create a compute cluster to support text analysis of
content in the cloud
10. Storage RFP
• Storage hardware RFP issued Dec. 20, 2013
• High density disk storage servers (DSS)
• Evaluation and analysis through early March
• Awarded to Dell: 2nd week of March
• All equipment delivered by 31st March, 2014
11. Data Storage Server
• PowerEdge R720xd
• MD1200 disk drawers
• Each drawer contains
48TB (12 x 4TB NL-
SAS drives)
• DSS capacity: 48TB to
432TB:
14. GTAnet Pilot
Purpose of the pilot:
To understand how to design and implement an effective
network topology to support the operation of the OLRC
Storage Cloud
15. GTAnet Pilot
Execution of the pilot:
Model and record the network traffic generated between
four OpenStack Swift storage nodes during routine
operation and under various simulated disaster
scenarios
18. Swift Node Considerations
How much bandwidth can they provide?
Will they enable jumbo frames?
Will they extend VLANs across their
network?
How low are their OTO & ongoing costs?
Do they have an ORION POP on site?
30. Status
Beta!
Develop end-user tools
Repository integration
Compute cluster and text mining
31. Acknowledgements
Our Partner Libraries
GTAnet – Doug Carson, Lloyd Kwong, Kevin Wong
ORION – Andy Lam, Mark Grant
OLRC Admin & Tech Committees
SP/UTL Systems teams – Steve Baroti, Chris Crebolder,
Miki Wong, Harpinder Singh, Bikram Singh
32. Interested in Learning More or
Getting Involved?
cloud@scholarsportal.info
https://spotdocs.scholarsportal.info/display/ODLRC
Notas del editor
The slide that will live in infamy.
logos
Develop a 1.2PB object storage service for partners
The majority of the monies awarded was for hardware
The first step was to buy the hardware. Under the rules of the grant, money had to be spent by March 31st.
Server has 2 H800 RAID controllers
Up to 4 MD 1200 directly attached to EACH controller in a redundant fashion (total of 8 per server)
RAID 0
GTAnet is the Regional R&E network for the Greater Toronto Area
3 Partner Libraries (Ryerson, York, UofT) are connected via GTAnet. We consulted with the folks at GTAnet and they said they’d be happy to help setup a network for our pilot. They were connected
Use VLAN technology over shared 10GB connection physically limited to 1GB running over GTAnet routers extended through backbones at Ryerson, York and Toronto
- Each site is assigned a private /24- VLANs are configured over existing GTAnet connectivity- routing is performed at the GTAnet router- RyersonU, UofT and YorkU extend VLANs through their backbones
IMPORTANT NOTE: On the graphs, ‘Inbound’ and ‘Outbound’ are reversed from what you would expect
Round 1 - Three storage nodes were set up, one each at UofT, Ryerson, and York. Additionally, a proxy node (which controls routing of traffic to the storage nodes) was set up at UofT. Once operational, baseline network traffic between the nodes was negligible (5-10Mb/s).
Round 2 - 35TB of data was loaded into the cluster. At 1Gb network speeds, replication and distribution of the data across the cluster took a couple of weeks (in actual time, it would be 9.72 days). Network utilization for the storage nodes averaged 350 Mb/s, but the proxy node's 1Gb network connection was fully utilized the entire time.
Round 3 - Once loading of the 35TB was completed, 12 drives were added to each storage node, which led to a rebalancing of the content within the cluster. Network utilization for the storage nodes hovered between 500-850 Mb/s, while the proxy node did not come under significant load. Attempts to load additional content at this time led to complete saturation of the 1 Gb network connections between the storage nodes.
Round 4 - This test added an additional storage node at the UofT site, which led to another rebalancing of the loaded content. It took approximately 2.5 days (58 hours) to import ~26.25 TB of content to the new node, during which time, the 1Gb connection to the new node was fully saturated. Network utilization of the other nodes was around 350-500 Mb/s.
It is worth noting that these scenarios represent extreme changes to the storage network. Outside of initial setup conditions, upgrades and other planned events will occur in a much smoother fashion, which should ameliorate the need for such high utilization. That said, they may be fairly representative of possible outage scenarios, in which the network might lose a node and have it reappear.
Typical use cases would be if one hard drive dies. Each hard drive is 4TB in size and over a 1Gb pipe, it would take 8.88 hours
ORION, GTAnet, and SP staff came up with this:
- 1 R720xd and 4 drawers as well as a switch will be installed at each site- ORION will provision a VPLS network to link the nodes
sites would pass the ORION VPLS VLAN through their networks to the OLRC switch
Under this architecture, each node can talk directly to each other without having to hit the GTAnet router.
I want to end with this chart that represents the amount of time it would take to rebalance our cluster (and bring it to 100% health) under our current configuration (which is 1 server with 4 drawers) if there was 200TB of data on the cluster.
On the right are “failure zones”
We anticipate that the 2 most common occurrences will be drives failing and Zones failing (because this could happen if network connectivity was lost to that zone, and on a WAN, the odds of this happening increase).
Drive | 4.444444444 | 2.222222222 | 1.481481481 | 1.111111111 | 0.888888889
Drawer | 53.33333333 | 26.66666667 | 17.77777778 | 13.33333333 | 10.66666667
RAID Card | 213.3333333 | 106.6666667 | 71.11111111 | 26.66666667 | 21.33333333
Zone Hours | 266.6666667 | 133.3333333 | 88.88888889 | 66.66666667 | 53.33333333
Zone DAYS | 11.11111111 | 5.555555556 | 3.703703704 | 2.777777778 | 2.222222222