Hi Raj, To summarize the requirements for everyone CC’ed your two objectives are: 1. Baseline testing a traditional DTN which accesses the shared filesystem over a network fabric, compared to a DTN directly on a storage controller. The JLSE file servers provide reasonably facsimiles for embedded storage controllers as they have a point-to-point DDR IB connection to the D2A9900 controllers. They’re also lower powered servers so should simulate the lower resources available on an embedded storage controller virtual machine. 2. Storage-aware Large Data Transfers: Using data locality information from GPFS to direct GridFTP to the DTN which has local block access to the data. Thus allowing GridFTP to be more efficient in its use of the storage network (Myrinet in this case). A good analogy is Hadoop’s HDFS, but for data transfer instead of compute. Please correct or expand the above as needed. Due to the Petrel testbed being setup as a single file system, and being a pseudo production testbed for Globus Share work (i.e. we can't just destroy the existing globus-fs0 and rebuild it), I think the best course of action to support the above work is to setup two 8 file-server filesystems on a new Myrinet fabric. We’ll initially create a file system on one (or optionally both) sets of 8 file servers. This will let you use opposing file servers for the baseline testing across a network fabric. This testbed would be similar in specs (same hardware, same networking) as the Petrel testbed. Once that testing is complete, we will recreate the GPFS as a single file system so you can experiment with data locality. Without further configuration this will give you two silos of data locality (two S2A9900 racks). If you want to further worsen the test cases, we can limit which LUNs are presented to which file server, lowering the number of file-servers with block access to a given set of data. It looks like we could have this testbed put together by early next week, as we have a near ready set of 16 file servers available. I’m also suggesting we use this new testbed as a more general resource for those that need GPFS but can withstand the file system being destroyed. I believe this fits Bill’s project needs, and he has similar goals in testing data locality with GPFS. All, please let me know if you have any comments or concerns. Thanks, Ben