IT Staff Convention 2008:Supporting Research Infrastructure

From Provider Notes

Supporting Research Infrastructure

Moderators: Michelle Gallert and Brandon Lodriguss, Wharton

Notes: Michelle Gallert, Wharton

What is infrastructure as defined by the group?

• Hardware
• Software
• People involved in creating in and requesting it
• Network
• Space
• Air Conditioning
• Power
• Policies
• Procedures

Centralized vs. Distributed

• In Wharton, there has been a recent initiative to switch back from decentralized servers to one large centralized server.
• SEAS, bad experiences (with centralized server) in the past with people in different groups, that don’t have shared goals and shared policies and shared research. It got very messy. Would say that servers should be allocated by user groups. Centralized within user groups, but decentralized between the various groups. At the time of using this, they did not have access to Condor – where you can achieve a higher level of priority and allocation than could have been used before.
• One Issue with allocating resources - People can bypass the system by directly logging into the grid engine. They had a hard time not allowing that feature.
• Summary of shared resources. Scheduling is central. You want a fair allocation of the shared resources. There are ways to program it into the system, or you can talk directly to the people and set policies.
• Different Software Packages: Torque, (great for talking to nodes all at the same time) Maui, (great for talking to nodes all at the same time) and Sun Grid System (great for batch processing), Condor (spare cycles – seeing when space is available when it is not being used) are some popular software packages to use. They are all for clusters. Sharing a bunch of computers at the same time. Each one has its advantages and disadvantages.
• Paying: some of the software packages have options for tracking usage, but it is not a great method.
• One system that was used within SEAS: for free, users could submit low priority jobs for free, and if the user wanted to submit high priority jobs, then you could pay to do that. Ask your users what they want to pay for and then talk about what hardware could be made available to them.
• There is also a trend that people want to purchase their own equipment, but then they realize that they do not know how to support it and they also have this powerful machine that is not being utilized a large percentage of the time. How can we tap into that unused resource?
• Instead of moving everyone to one system, could you make it easier for the decentralized people to support these systems?
• Standardization (the build) was mentioned several times when talking about having separate machines.
• All equipment is Centrally managed, but they are charged according to how far they leave the standard build
• Sun Grid Engine – you can give people sudo root rights that allow administration at a lower level
• What about individual servers that people can go out and buy at the same price for renting off of the centralized server? To combat this is to offer a big enough carrot to moving to centralized – managed, backed up, power and cooling control…

• Data Storage: Run out of data storage room, looking to centralize among the university by paying other parties instead of adding on to their own storage.
• Backups – tape vs. disc to disc space. Don’t have the budget to do their own budgets.
• Centralized storage: ask for a grant to be able to backup and have storage.
• ISC does offer storage space that is backed up but for a very high price. SEAS merges some groups together to get back some of their sunk costs that they bought for themselves and give it to others. The problem is that many research needs are different as far as backups.
• Pooled money allowed one LSP to provide backups and a SAN in his department. But money is a big issue.
• Another model that could be cheaper to implement - every department buys its own machine that they are backing up, while also utilizing the central software and hardware as the 2nd tier of backup.
• On a grid utilized by several departments, who is the first tier support for the system – power users within the department. Each research group has their own methods of how they can link with the centralized server
• Mailing list has already been created that can be used.
• ISC provider WIKI, there are tags for these topics
• How do you decide who to buy from? Some companies offer evaluation units

Link to article within Provider Notes IT Staff Convention 2008: Session Overviews

Link to article on a website external to Provider Notes External Link