The Large Hadron Collider
Once the decision was made to decentralize the analysis resources, a crucial question needed to be answered. How does a physicist in Europe run a job using data stored at Nebraska? In 2004, the computing model for CMS was finalized and embraced the emerging grid technology. At that time, the technical implementation was left flexible to allow for sites to adopt any grid middleware that might emerge. Analysis sites in Europe adopted the World LHC Computing Grid (WLCG) software stack to facilitate analysis. Sites in the US chose the Open Science Grid (OSG) to provide the software to deploy jobs remotely. The two solutions are interoperable.
The OSG's (www.opensciencegrid.org) mission is to help in the sharing of computing resources. Virtual Organizations (VOs) can participate in the OSG by providing computing resources and utilizing the extra computing resources provided by other VOs. During the past year, the OSG has provided 280 million hours of computing time to participating VOs. Figure 1 shows the breakdown of those hours by VO during the past year. (The Internet search for the meaning of the VO acronyms is an exercise left to the reader.) Forty million of those hours were provided to VOs not associated with particle physics. Participation in the OSG allows Nebraska to share any idle CPU cycles with other scientists. Furthermore, the CMS operational model for all US Tier-2 sites is that 20% of our average computing is set aside for use by non-CMS VOs. This gives non-CMS VOs an incentive to join the OSG. Non-CMS VO participation increases support and development of the OSG software that allows CMS to benefit from improvements made by other users. The OSG's model should serve as an example for similar collaborative efforts.
OSG provides centralized packaging and support for open-source grid middleware. The OSG also gives administrators easy installation of certificate authority credentials. Certification and authentication management is one of the OSG's most useful toolsets. Further, the OSG monitors sites and manages a ticketing system to alert administrators of problems. Full accounting of site utilization is made available by OSG so that funding agencies and top-level management have the tools they need to argue for further expenditures. See Figure 2 for CPU hours provided to the OSG by some of the major facilities.
In short, what SETI@home does with people's desktops, OSG does for research using university computing centers.
The CMS experiment will generate more than one terabyte of recorded data every day. Every Tier-2 site is expected to store hundreds of terabytes on-site for analysis. How do you effectively store hundreds of terabytes and allow for analysis from grid-submitted jobs?
When we started building the CMS Tier-2 at Nebraska, the answer was a software package written at the high-energy physics (HEP) experiment DESY in Germany called dCache. dCache, or Disk Cache, was a distributed filesystem created by physicists to act as a front end to a large tape storage. This model fit well with the established practices of high-energy physicists. The HEP community had been using tapes to store data for decades. We are experts at utilizing tape. dCache was designed to stage data from slow tapes to fast disks without users having to know anything about tape access. Until recently, dCache used software called PNFS (Perfectly Normal File System, not to be confused with Parallel NFS) to present the dCache filesystem in a POSIX-like way but not quite in a POSIX-compliant way. Data stored in dCache had to be accessed using dCache-specific protocols or via grid interfaces. Because file access and control was not truly POSIX-compliant, management of the system could be problematic for non-dCache experts.
dCache storage is file-based. All files stored on disk correspond to files in the PNFS namespace. Resilience is managed via a replica manager that attempts to store a single file on multiple storage pools. Although a file-based distributed storage system was easy to manage and manually repair for non-experts using dCache, the architecture could lead to highly unbalanced loads on storage servers. If a large number of jobs were requesting the same file, a single storage server easily could become overworked while the remaining servers were relatively idle.
Our internal studies with dCache found that we were having a better overall experience when using large disk vaults rather than when using hard drives in our cluster worker nodes. This created a problem meeting our storage requirements within our budget. It is much cheaper to purchase hard drives and deploy them in worker nodes than to buy large disk vaults. The CMS computing model does not allow funding for large tape storage at the Tier-2 sites. Data archives are maintained at the Tier-0 and Tier-1 levels. This means the real strength of dCache is not being exploited at the Tier-2 sites.
The problems of scalability and budgeting prompted Nebraska to look to the Open Source world for a different solution. We found Hadoop and HDFS.
Hadoop (hadoop.apache.org) is a software framework for distributed computing. It is a top-level Apache project and actively supported by many commercial interests. We were not interested in the computational packages in Hadoop, but we were very interested in HDFS, which is the distributed filesystem that Hadoop provides. HDFS allowed us to utilize the available hard drive slots in the worker nodes in our cluster easily. The initial installation of HDFS cost us nothing more than the hard drives themselves. HDFS also has proven to be easy to manage and maintain.
The only development needed on our end to make HDFS suitable for our needs was to extend the gridftp software to be HDFS-aware. Analysis jobs are able to access data in HDFS via FUSE mounts. There is continued development on the analysis software to make it HDFS-aware and further remove unnecessary overhead.
HDFS is a block-based distributed filesystem. This means any file that is stored in HDFS is broken into data blocks of a configurable size. These individual blocks then can be stored on any HDFS storage node. The probability of having a “hot” data server that is serving data to the entire cluster starts to approach zero as the files become distributed over all the worker nodes. HDFS also recognizes when data required by the current node is located on that node and it does not initiate a network transfer to itself.
The block replication mechanisms in HDFS are very mature. HDFS gives us excellent data resiliency. Block replication levels are easily configured at the filesystem level, but also can be specified at the user level. This allows us to tweak replication levels in an intelligent way to ensure simulated data that is created at Nebraska enjoys higher fault tolerance than data we can readily retransfer from other sites. This maximizes our available storage space while maintaining high availability.
HDFS was a perfect fit for our Tier-2.