Hi Andrew and the OGSA-WG, I apologize for missing the meeting last Thursday on this topic. We have a machine full of new cluster and grid equipment, and I have been fully occupied commissioning and configuring it. I am afraid that I differ rather strongly with the direction being taken with regard to the HPC profile at this stage. My view is strongly that simple username/password login, even SSL secured, is quite demonstrably insufficiently secure to deploy as a model for authentication and access to high performance computing. I disagree fairly strongly that any sort of stop-gap of this nature should be written into the HPC profile, distributed or promoted at this time. I have an excuse for having taken so long to reply on this topic. It was necessary for me to investigate as thoroughly as possible the current state of deployment of GSI-secured alternatives to username/ password login and to do so in a way that would allow me to give a credible response to all of you regarding the state of the art on this topic. At this point I am assured and feel sufficiently confident to proceed, either at OGF-19 or before, with Andrew, Marty, and whoever else would like to participate on a revision of the HPC profile that would cover more secure basic access to high performance cluster and storage systems based on GSIOpenSSH and similar software that uses either GT4 or an equivalent callout. We are writing standards, not implementations, but I wished to be sure that the state of the art on existing implementations would be consistent with making this recommendation. It is essential from my point of view to promote secure access to HPC resources. As the bulk of the compromise attacks that have been successful over the past 2 to 3 years on HPC resources has been through discovery and reuse of username/password combinations from ordinary users (at least as I read the recent record), I think that now is not the right time to propose backing off from the use of strong cryptographic methods to use HPC resources in grid settings. The use of strong cryptography does not have to be limited to X.509 "pure classic" PKI, and I look forward to an active discussion on federated identity and related topics to be held at the OGF meeting next week. It is clear to me that recent improvements to the availability and technology for authentication, authorization and attribute transmission will make many modes of access to grid resources possible with appropriate security that up to now have been either impossible or confined to limited implementation. For the moment, I would like to suggest that a revision of the HPC profile propose that "only GSI or equivalently secure architectures be used for direct access to HPC resources" and that the document be revised specifically to discourage the direct access by users to highly capable computational and to secure storage resources by username/password mechanisms. In my own project, we use GSI-OpenSSH via grid-mapfiles. I have been able to confirm that current implementations of GSI-OpenSSH are capable of interoperating with more general callout-based systems, including attribute-based AuthZ systems, without modification. Therefore it is not necessary for users to have username/password access if direct login is needed on an HPC system. As a further enhancement to the document and to the profile, I feel it would be useful to describe architectures for pure-computational (i.e., batch-only access), for pure-login (i.e., front-end and submission access), pure-storage (i.e., stage-in/stage-out and related data handling) and for the interesting use case of "managed fork" (i.e., interactive but sand-boxed grid access) systems. I believe these changes would result in an improved HPC profile that would be of better total usability within the HPC community. This document is NOT attached, instead your original one is for discussion, but I believe can be worked out in the context of discussions to be held at OGF-19 next week. Sorry for being (apparently but not really) strident, but I believe the above reflects current best practices better than recommending username/password support for direct login to HPC systems. I would not personally be able to support the current draft as written. Thanks and best wishes, Alan On Jan 18, 2007, at 2:16 PM, Andrew Grimshaw wrote:
All,
On this mornings call I volunteered to see what was up with the HPC profile working group with respect to authentication. Recall that we need some sort of authentication story in the short run or we cannot put together any form or realistic, cross-organizational, compute grids with BES, or for that matter data grids using RNS/ ByteIO.
Attached is a short white paper from the HPC Profile WG (or maybe just the three authors). It is BES-specific, but I think the ideas may be generalized to a broader set of OGSA services. I think we should consider it, or something like it.
Note that it does NOT deal with the ultimate authentication and delegation problem that we will face. Rather, I personally (speaking only for myself, and not even the people in my research group) think that this sort of solution is a stop gap that we can use for awhile, and that we will ultimately deprecate in favor of whatever comes out of the OGSA-Authentication WG.
So, for your reading pleasure – and with my thanks to Marty for giving me a copy.
A
Andrew Grimshaw
Professor of Computer Science
University of Virginia
434-982-2204
grimshaw@cs.virginia.edu
--
ogsa-wg mailing list ogsa-wg@ogf.org http://www.ogf.org/mailman/listinfo/ogsa-wg
Alan Sill, Ph.D TIGRE Senior Scientist, High Performance Computing Center Adjunct Professor of Physics TTU ==================================================================== : Alan Sill, Texas Tech University Office: Admin 233, MS 4-1167 : : e-mail: Alan.Sill@ttu.edu ph. 806-742-4350 fax 806-742-4358 : ====================================================================