Here are three press releases about three very different branches of science going on at SLAC.
The first is about our biggest experiment, BaBar. They proved a certain kind of CP ("charge-parity") violation that explains in part the asymmetry between matter and antimatter. This explains (in part) why after the Big Bang more matter than antimatter was left over and the universe could form. BaBar has taken around 2 PetaByte of data sitting on tape but all their development tools and software code sit in AFS. BaBar uses AFS forworldwide collaboration, and distribution of their code repositories.
September 28, 2006 - New Form of CP Violation Discovered
The second one is astrophysics research done at KIPAC (Kavli Institute for Particle Astrophysics and Cosmology). They did mass analysis of pictures from the Chandra X-ray Observatory and for the first time directly observed the Dark Matter. The data itself is again not in AFS but most of the development tools and software code is in AFS.
August 21, 2006 - Dark Matter Observed
The next two are about our Synchrotron Radiation Lab (SSRL). They highlight the application of SLAC's physics technology for completely different branches of science.
August 2, 2006 - Modern Technology Reveals Ancient Science
Archimedes Manuscript Yields Secrets Under X-Ray Gaze
SLAC is currently expanding the computer center's relationship with SSRL to provide AFS space for their tools like Gaussian (sequential and parallel). They have also started to use our standard software development tools (compilers, etc.) in AFS space.
Lastly, a very widely used toolkit, GEANT4 (http://geant4.web.cern.ch/geant4/), for the simulation of the passage of particles through matter is being co-developed at SLAC. Everything they do is in AFS. Among many, many other things, this toolkit is used for medical simulations of radiation treatments of tumors in the human body.
A summary of research performed at SLAC in 2006: http://today.slac.stanford.edu/feature/holiday-message06.asp
Since 1998, SLAC has stored all user home directories in AFS. AFS is available to the several thousandmachines in its batch farm and is used primarily for the delivery of binaries for batch jobs.
How much data (in TB)? 5.5TB
How many users? 3247
How many volumes? 18621
How many file servers? 15 fileservers
How much storage per file server? 34GB to 1.6TB
Reasons important back then but not as important any more:
Reasons still important:
Challenges for the future:
AFS usage at PDC in numbers (2007-01-01):
Number of Clients: Unknown. PDC manages 5 clusters with a total of more than 1200 computing nodes. When the clusters are new, they use to appear around rank 70 of the top 500 list. All production systems use AFS. There may be a single digit number of computers which have no access to the AFS space. Add to that the number of researchers who access their data from their own workstations from all over the Internet.
Number of simultaneous users: 100-200.
Number of users with $HOME in pdc.kth.se AFS space: Approx 3000.
Amount of data in AFS: Approx 5TB.
Number of volumes: Approx 6000.
Number of files: Approx 65 000 000.
Number of file servers: 11
Approx storage per server: 1TB
Pictage's business is worldwide. They provide services to wedding photographers who upload the photos they have taken. Pictage then makes the photos available to the wedding party for photo selection and printing. Photos selected for printing are retouched by hand using Adobe Photoshop.
Pictage uses OpenAFS to store all of the original photos, the processed photos, and the web content.
Work flow is managed by the use of mount points and symlinks which are placed within directories assigned to various photographers and their associated events.
Pictage storage has exceeded 200TB of data and more then 40 million files. They frequently push the limits of the AFS directory structures running into the directory entry limitations.
Directory search performance is one of their bottlenecks. Unlike more modern file systems, AFS does not use a B+ tree representation and directory searches within the case-insensitive Microsoft Windows client are linear in nature. This wastes significant CPU utilization and clock time. Pictage had measured the performance loss on a duo-core Xeon processor at between 3 and 5 seconds to open a file in Adobe Photoshop when the directory contains 15,000 entries, because Adobe Photoshop will request information forthe last 30 files it has seen, thus requiring heavy usage of AFS directories. OpenAFS for Windows 1.5.13 added an optimization for cached files with registered callbacks that avoids the performance problem for files that are known to exist. With this optimization in place, Pictage experienced file open times similar to those experienced when opening files on the local filesystem. For files that do not exist such as Java class files, the optimization does not apply, and the linear search must still be used. Number of simultaneous users: 150 AFS client users; 1000 web server users
Amount of data in AFS: Approx 215TB
Amount of AFS storage: 265TB with planned growth to 425TB in twelve months
Number of volumes: Approx 800,000.
Number of files: Approx 200 000 000.
Number of file servers: 70 with 10 more planned
Approx storage per server: depending on age of server 1.2TB to 25TB
(statistics as of 22 May 2007)
HDF is a project that helps support the research projects that NCSA provides computing services for.
This page is prettier and has some more concise history:
They've recently split off from NCSA to form The HDF Group:
Their source tree and everything is in the NCSA AFS cell and the build process is scripted for all the different architectures they support. Their build systems range from clusters to servers to desktop systems and the binaries can be dropped right into AFS which users grab via ftp. They need an authenticated file system that supports all of those platforms.
NCSA's Security team uses AFS for storing security incidents and collaborating with off-site users.
And some comments from Christopher Lindsey about future directions:
"I can tell you about possible future direction.
"We've all noticed that the stand-alone University model doesn't work. Just like businesses that merge, Universities are finding that they need
to collaborate with other research institutions to make their proposals competitive. You throw in concepts like grid computing (i.e. TeraGrid)
and you suddenly have a diverse, disparate set of sites that need to collaborate.
"This is where AFS becomes a player.
"I believe that we currently use it as a giant swap file space for different projects -- one supercomputer site needs access to security
net flows, so we put them out there, set the acls, and voila.
"Future projects on supercomputers will leverage AFS as well. I don't know how much detail I can go into, but there is a proposal that will
be submitted on Jan 22 that will revolutionize how HPC resources within the TeraGrid are used. The ubiquity and security of AFS make it the
perfect foundation for this project.
"As far as timings, we might be able to do some dprof timings between here and Argonne Labs over iwire (http://www.iwire.org/). Ironically, the bottleneck there will be the GigE NIC on the AFS servers."
The use of AFS at the United States Geological Survey, USGS, started in response to a requirement to provide natural hazard information on the Web so that it could continue being delivered to the public even in the event of large-scale regional disasters. The driver for this activity was the inability of US Geological Survey to provide water level measurements during hurricane Floyd in 1999 just when that information was needed most by emergency managers, because our database servers and Web servers were rendered useless by power and networking failures in affected states.
This regionally distributed, "no single point of failure" delivery Web service relies in part on AFS to provide regionally replicated file content at "modules" housing AFS file servers and Web servers in Menlo Park, California, Sioux Falls, South Dakota, and Reston, Virginia.
AFS was chosen as it was the only multi-platform software available to provide geographically distributed content replication, access control, and user transparency as required by our project.
Three Web servers at each module have an identical view of read-only copies of data provided by AFS. Bringing on additional Web serving capacity is as easy as adding another Web server which is also an AFS client. A commercial DNS service with wellness checking, Akamai, resolves hostnames to these geographically distributed Web servers, which are checked every minute for availability. Failure to contact Web service on any one of these systems results in the corresponding IP address being taken out of DNS resolution.
This National Web server system known as, NatWeb, delivers Web content for 150 unique USGS websites. In 2006 NatWeb delivered an average of 1.2 million web pages and 191 gigabytes of data per day, for an annual total of 453 million pages and 68 terabytes. It has experienced a roughly 50% increase in utilization annually for the last three years. We serve data as diverse as
Our use of AFS is tied nearly exclusively to Web service. All content is replicated to all regions. The R/W master for content is located in the same region as content maintainers, so R/W and replicated R/O information is distributed evenly across regions. Any one region can drop from the network, and the associated Web information will continue to be served to the public, although updates cannot be accomplished in this condition without administrator intervention.
We are a small AFS site by most measures. AFS usage specifics (with some rounding):
Our architecture uses custom programming to detect changes to AFS content. Volumes containing updated content are automatically released (replicated to remote locations) every 15 minutes. The file servers collectively execute over 2000 "vos release"s per day.
Our server implementation is entirely Sun/Solaris 10 based. At each of our three module locations:
There are three hostname/IP address pairings by which all AFS file services are known, as well a separate hostname and IP address which is always associated with the same computer. By shutting down AFS services, and shuttling the appropriate IP address from one system to another, we can effectively fail over AFS file service from one system to another.
Our AFS clients fall into two major categories, the Web servers which serve content to the public, and the content maintainers who have AFS installed on their desktops.
Our content maintainers are overwhelmingly using Windows XP, with perhaps 5% using Mac OS X and still smaller numbers using Linux and Solaris.
Our entire staff consists of 3 full-time employees, 1 part-time employee, and one manager. These individuals manage the budget, hardware, operating system maintenance, security, software updates, custom programming and user support for both the AFS and Web components of the service.
We're using the OpenAFS client code pretty extensively, and have been since fairly early in its release cycle. We're running the OpenAFS client on all of our Windows lab machines (we cheered the day the Windows clients became available, since we'd never found the funding to put Transarc clients on our Windows machines, and since the students had been clammoring for access to their AFS home directories for quite a while). We're also using the OpenAFS client on our MacOS X systems, although it will be summer before we actually roll out OS X in our campus lab environments. We've recently started a publicity campaign on campus to get students and faculty to take advantage of the OpenAFS client installation on their personal computers wherever possible, viewing AFS access from desktop systems throughout the campus as one of the primary ways to move toward decomissioning insecure FTP access to our central facilities -- we find that running AFS clients on desktop systems greatly simplifies our users' use of web development tools like DreamWeaver, since with the client, they can publish changes directly into their AFS home directories.
On the whole, we've had very few problems with the client -- most of them related to the installer or to versioning issues with earlier versions of the code. We're still working on completing our integration work with Windows 2000 and XP -- we've established cross-realm trust between our nascent Active Directory and our existing K5 realm, and we've been working on trying to arrange to have AFS tokens derived locally from the K5 tickets granted to users on our Windows systems at login time. So far, that's not worked quite as well as we'd hoped, but that's purely a Windows problem, not an OpenAFS problem -- we've worked around it with some simple scripting at the client end for now, but I'm hopeful that when we get a chance to do some more developement work over the summer, we may find a means to make tokenization transparent on our Win2K/XP systems.
We're hoping to do some testing early in the summer with the OpenAFS server, and provided that we can get everything to work out, to start migrating our server farm from the Transarc server code to the OpenAFS server code. We've been using AFS at our site since the early 1990s and we have a modestly-large AFS cell comprising Solaris servers of various vintage (Solaris 2.6, mostly, with some Solaris 8 on our more recent acquisitions).
Our cell runs on OpenAFS fileservers on Tru64 5.0a. We have 0.5 - 1 TB of data and about 250-400 clients. We are VERY pleased with the fileserver, it's quite stable, no crashes in the 6 months we've been running them.
Our 15 Solaris clients are running OpenAFS, we are very pleased with this one. (We plan to have about 100 Solaris hosts in a few months.)
I first used AFS in 1995 during my first year as a student. Now I am a member of the academic staff at Chemnitz University of Technology. Two years ago I started helping the Dr. Wilhelm Andre Gymnasium (Gymnasium is the German word for high school) to build up their computer lab. At that time we had only 30 workstations and one server using NFS.
Last year (2001) we were very lucky and got enough money to build up a large intranet. At least large for a school. We now have 4 fileservers (each with 400 GB hardware RAID) with a total of 1.6 TB AFS space. Normal school-installations have to live with 50 to 100 GB on Windows NT or Novell.
We have 1500 users and 100 clients (the next 50 clients are in sight). The primary platform is Linux. Some workstations run Windows 2000 for the sake of legacy applications.
Before OpenAFS, we used NFS and Samba. There were no real problems with that solution because we had very little filespace (20 GB) on one server. But we saw the problems arising with 4 servers and much more capacity: The load on the server would rise and there was no transparent way to reorganize the location of the directories without heavy work on the users side due to absolute path names in configuration files! Neither NFS nor Samba provide ACLs in a smooth way. It would have required patching and installing lots of additional software. And quotas are not really easy to use. And last but not least: AFS is just cool :-)
OpenAFS saves us a lot of time. This starts with the installation of our workstations via RedHat KickStart and NFS. Installing 100 clients in parallel (each requiring about 1.5 GB of binaries) produced an insustainable load on one server. So we decided to create an install volume and have it replicated on each AFS server. Each AFS Server machine runs a NFS re-exporter. With a round robin DNS-entry for the NFS installation server we can quadruple the capacity of our install source without writing cryptic KickStart configurations.
Of course all the well known benefits of AFS (automatic backups, fault tolerance, etc.) help us save a lot of time and stress.
One of the school specific problems with AFS is that we need an operation mode for our workstations and servers allowing users to log in using AFS authentication, but without allowing access to the home volumes of the users. We need this mode for writing "electronic tests". We try to solve this with negative ACLs for a special test-writing-group.
The next problem is to distribute some of the rights of system:adminstrators without making users members of that group. For instance we like to give teachers the right to change quotas of their puplis' volumes but not their own. We think we will solve this with a middle-layer web page.
At the moment we do not buy any commercial support. But we are trying to extend our cell through the whole city with 160 sites and 50000 users. At that time we will need some help.
Since writing this story two years passed by.
Currently we have 1733 volumes on-line, 1585 of them user-volumes. The aggregated quota is 210GB with 90GB used. The initial user-qouta is 50MB. With the help of a PHP-based web-tool the qouta can be changed by a teacher. Basic fs- and vos-operations (setacl, mkmount, release) are done by this teacher manually at command line-shell.
In this 2 years of operation there was not a single-data-loss because of AFS (not that there has been any data-loss at all :-) Even the sudden death of a file-server-machine didn't corrupt data. The second incident of this kind wasn't event noticed for some time.
The 4 AFS-server-machines (Athlon 1400, 1GB RAM, E1000-Network) still have an avg-load of 0.02. Obviously there is no need in replacing them in near future.
3 Months ago we performed an update from RedHat Linux 7.2 to Fedora Core 1. Except some minor troubles with the NFS-reexporters and the cfengine it worked seamlessly. At the same time OpenAFS 1.2.11 was installed without any pain at all. OpenAFS 1.2.3 was used before.
We use our file-servers not only for user-files but also for backup and system-maintainance files. Daily system-backup of all servers (mail, ldap, ...) is done to an AFS-volume for each server. The volume is replicated to another fileserver. We use cfengine for administrative tasks. The cfengine-repository is stored in AFS and distributed by AFS and not the cfengine-file-server.
The balancer is used to spread load between fileservers by moving volumes in respect of week-usage and number of volumes.
As a next step we will add an fifth fileserver on a remote location (only connected with 128 kbit). We hope that by replicating a ro-db-server and many volumes to that server will be sufficient to keep things working.
Computing Services rolled out OpenAFS to both clients and servers early, and now runs modern versions on Solaris 2.6 servers, as well as on Solaris 7 and 8 and locally-updated Redhat 6.1 clients. Additionally, the OpenAFS Windows NT client is being tested general rollout.
Despite often using pre-release versions of the software few problems have been experienced, and fixes have generally been quickly forthcoming, either internally done and contributed back to the community, or thanks to the many active contributors to the ongoing success of the project.
University Information Systems (UIS) started testing OpenAFS about 7 months ago for the purpose of binary standardization across our web, application, and database servers. Those 7 months have been very stable and we are starting to roll our cell into production. We currently have ~140gb of vicep space across four servers (2 linux, 2 solaris8). The cell gracefully handled two events of failed disks without interruption.
OpenAFS has and will continue to help me keep our 45 web, application, and database servers running the same versions of binaries and scripts.
As our programmers learn of the features and capabilities of OpenAFS they increasingly want their workstations added as clients also (mixture of sun and linux boxes). As trust builds in OpenAFS we will also start moving production data into volumes rather than local storage. This will allow applications to "float" between nodes as necessary.
Our site has been a long time Transarc customer for our academic cell (2.6Tb, 6 fileservers, 17000 users, 100s of workstations), The OpenAFS project and its contributors are breathing new life into a wonderful technology.
Some features that would be of value:
I am working on converting all my various perl scripts into a perl module so i can contribute it and open it up for community development.