70 likes | 163 Views
HPC USER FORUM I/O PANEL April 2009 Roanoke, VA. Panel questions: 1 response per question Limit length to 1 slide. Panel Format. <insert panel format here>. Q1. Parallel NFS finally is here!.
E N D
HPC USER FORUM I/O PANELApril 2009Roanoke, VA Panel questions: 1 response per question Limit length to 1 slide
Panel Format • <insert panel format here>
Q1. Parallel NFS finally is here! • With the formalization of Parallel NFS as a standard - what steps are being provided to enable this to be hosted on current (and future) platform choices? • This is really a development question for Garth and Peter • Consumers should require this as part of the RFP process • Remember that pNFS is a protocol and not a file system • POSIX semantics for writing to the same file are not maintained • Though pNFS is a very good thing is does not solve world peace and global warming • The underlying file system might need to be modified to support the new high speed I/O
Q2. Parallel NFS – implementation details… • What tools are available to help optimize this (from application layer level all the way to archival stage)? What is missing and who should provide it? • Actually there are no tools for end-to-end integrated I/O optimization and it would require a major effort to provide an integrated tool kit • There are many obstacles to these types of tools such as understanding the configuration and contention from other processes • Example: A user could make a 30% performance improvement I/O of a job and it might not be recognized given contention from other work • How do you measure the I/O of a job without measurement of the system
Q3. Physical media interconnects … • We all are facing complexity and cost issues. With IB or 10 GbE (40/100 GbE) : where should the HPC community should focus its resources on - for all I/O? • The community should be looking to the future at what will be the commodity technology will be the winner • That in my opinion will be FCoE and GbE • In my opinion sites should begin to plan for this future by planning for this by looking at NICs and cabling plans • The process of change will being in earnest in 2010 and given legacy environments will likely take at least 5 years to complete • Things take longer than expected with storage • Take the 3.5 to 2.5 disk drive change now in progress for an example
Q4. Layer protocols above the interconnects • Too many standards. interconnects, media layers are issues today. iSCSI/FCOE/FcoCEE/FCoIB have all been touted as the solution(s). Is it even relevant in the HPC arena? Is fragmentation the only choice? • In my opinion ISCSI will not make it as an HPC solution • The issue is encapsulation of metadata and small blocks and wasting channel bandwidth • HPC is not all about streaming I/O • Small block random happens more than you think and is the bounding factor for many HPC • Commodity technology will eventually overtake • IB was successful when PCIe became available • And it will limit us again • 40 GbE will win in cost over QDR • Maybe not at first but eventually given volume
Q5. I/O issues not yet addressed? • What do you consider to be the top 3 main (technical or human) issues in HPC I/O? • 1 HPC file systems have historically had a problem opening a single file for all nodes, which means the file counts are growing significantly • 2 There are no really I/O tools for users to tune applications on operational systems especially • 3 User data management (ILM) frameworks are not part of the standard process and are implemented in user space • Part of the complexity of this is a function of #1