From UABgrid Documentation
m (→Benchmarks: added NAMD benchmark link)
|Line 100:||Line 100:|
=== Performance Statistics ===
=== Performance Statistics ===
Revision as of 09:19, 22 June 2011
Cheaha is a campus resource dedicated to enhancing research computing productivity at UAB. Cheaha is managed by UAB Information Technology's Research Computing group (UAB ITRC) and is available to members of the UAB community in need of increased computational capacity. Cheaha supports high-performance computing (HPC) and high-throughput computing (HTPC) paradigms.
Cheaha provides users with a traditional command-line interactive environment with access to many scientific tools that can leverage its dedicated pool of local compute resources. The local compute pool provides access to three generations of compute hardware based on the x86-64 64-bit architecture. It includes 576 2.66 GHz cores of the latest generation Intel Westmere CPUs with a QDR Infiniband interconnect, 192 3.0GHz cores with a DDR Infiniband interconnect, and 120 1.6GHz cores with a traditional 1Gbs Ethernet interconnect. The compute nodes are backed by 180TB of high performance storage and an additional 40TB available for home directories and ancillary services. The compute nodes combine to provide nearly 3TFlops of dedicated computing power.
Cheaha is composed of resources that span data centers located in the UAB Shared Computing facility in the Business Engineering Complex (BEC) and the RUST Computer Center. Resource design and development is lead by UAB IT Research Computing in open collaboration with community members. Operational support is provided by UAB IT's Research Computing group.
Cheaha is named in honor of Cheaha Mountain, the highest peak in the state of Alabama. Cheaha is a popular destination whose summit offers clear vistas of the surrounding landscape. (Cheaha Mountain photo-streams on Flikr and Picasa).
In 2002 UAB was awarded an infrastructure development grant through the NSF EPsCoR program. This led to the 2005 acquisition of a 64 node compute cluster with two AMD Opteron 242 1.6Ghz CPUs per node (128 total cores). This cluster was named Cheaha. Cheaha expanded the compute capacity available at UAB and was the first general-access resource for the community. It lead to expanded roles for UAB IT in research computing support through the development of the UAB Shared HPC Facility in BEC and provided further engagement in Globus-based grid computing resource development on campus via UABgrid and regionally via SURAgrid.
In 2008, money was allocated by UAB IT for hardware upgrades which lead to the acquisition of an additional 192 cores based on a Dell clustering solution with Intel Quad-Core E5450 3.0Ghz CPU in August of 2008. This uprade migrated Cheaha's core infrastructure to the Dell blade clustering solution. It provided a 3 fold increase in processor density over the original hardware and enables more computing power to be located in the same physical space with room for expansion, an important consideration in light of the continued growth in processing demand. This hardware represented a major technology upgrade that included space for additional expansion to address over-all capacity demand and enable resource reservation.
The 2008 upgrade began a continuous resource improvement plan that includes a phased development approach for Cheaha with on-going increases in capacity and feature enhancements being brought into production via an open community process.
Software improvements rolled into the 2008 upgrade included grid computing services to access distributed compute resources and orchestrate jobs using the GridWay meta-scheduler. An initial 10Gigabit Ethernet link establishing the UABgrid Research Network was designed to supports high speed data transfers between clusters connected to this network.
In 2009, annual investment funds were directed toward establishing a fully connected dual data rate Infiniband network between the compute nodes added in 2008 and laying the foundation for a research storage system with a 60TB DDN storage system accessed via the Lustre distributed file system. The Infiniband and storage fabrics were designed to support significant increases in research data sets and their associate analytical demand.
In 2010, UAB was awarded an NIH Small Instrumentation Grant (SIG) to further increase analytical and storage capacity. The grant funds were combined with the annual investment funds adding 576 cores (48 nodes) based on the Intel Westmere 2.66 GHz CPU, a quad data rate Infiniband fabric with 32 uplinks, an additional 120 TB of storage for the DDN fabric, and additional hardware to improve reliability. Additional improvements to the research compute platform involved extending the UAB Research Network to link the BEC and RUST data centers, adding 20TB of user and ancillary services storage.
Grant and Publication Resources
The following description may prove useful in summarizing the services available via Cheaha. If you are using Cheaha for grant funded research please send information about your grant (funding source and grant number), a statement of intent for the research project and a list of the applications you are using to UAB IT Research Computing. If you are using Cheaha for exploratory research, please send a similar note on your research interest. Finally, any publications that rely on computations performed on Cheaha should include a statement acknowledging the use of UAB Research Computing facilities in your research, see the suggested example below. Please note, you're acknowledgment may also need to include an addition statement acknowledging grant-funded hardware.
Description of Cheaha for Grants
UAB IT Research Computing maintains high performance compute and storage resources for investigators. The Cheaha compute cluster includes 192 3.0GHz Intel-based compute cores with 386GB of RAM interconnected via a DDR Infiniband network. A high-performance, 60TB Lustre parallel file system built on a Direct Data Network (DDN) hardware platform is also connected to these cores via the Infiniband fabric. An additional 40TB of traditional shared storage and an auxiliary 120 1.6GHz AMD-based compute cores are available via a 1GigE network fabric. This core general access compute fabric 312 nodes and 100 TB of storage is available to all UAB investigators.
NIH funded investigators are granted priority access to an expanded compute and storage pool that includes an additional 576 3.0GHz Intel-based compute cores, 2.3TB RAM and 120TB in the high-performance Lustre parallel file system all interconnected via a QDR Infiniband network fabric. ScaleMP software is available on this enhanced compute fabric to combine these resources into aggregate system images that can leverage the enter collection within a single system image.
Acknowledgment in Publications
This work was supported in part by the research computing resources acquired and managed by UAB IT Research Computing. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the University of Alabama at Birmingham.
NIH funded researchers who leverage the additional priority resources available to them should also include a statement of support regarding this NIH funded hardware.
The Cheaha Compute Platform is composed of three generations of commodity compute hardware. Using the generation designators gen1, gen2, and gen3 described below is the clearest way to communicate about specific hardware collections.
- Generation 1 (gen1) -- 64 2-CPU AMD 1.6 GHz compute nodes with Gigabit interconnect. This is the original hardware collection purchased with NSF EPSCoR funds in 2005, approx $150K. These nodes are sometimes called the "Verari" nodes. These nodes are tagged as "verari-compute-#-#" in the ROCKS naming convention.
- Generation 2 (gen2) -- 24 2x4 core (196 cores total) Intel 3.0 GHz Intel compute nodes with dual data rate Infiniband interconnect and the initial high-perf storage implementation using 60TB DDN. This is the hardware collection purchased exclusively with the annual VPIT funds allocation, approx $150K/yr for the 2008 and 2009 fiscal years. These nodes are sometimes confusingly called "cheaha2" or "cheaha" nodes. These nodes are tagged as "cheaha-compute-#-#" in the ROCKS naming convention.
- Generation 3 (gen3) -- 48 2x6 core (576 cores total) 2.66 GHz Intel compute nodes with quad data rate Infiniband, ScaleMP, and the high-perf storage build-out for capacity and redundancy with 120TB DDN. This is the hardware collection purchased with a combination of the NIH SIG funds and some of the 2010 annual VPIT investment. These nodes were given the code name "sipsey" and tagged as such in the node naming for the queue system. These nodes are tagged as "sipsey-compute-#-#" in the ROCKS naming convention.
Summarized, Cheaha's compute pool includes:
- gen3 is 576 cores of 2.66GHz with 4Gb RAM per core
- gen2 is 192 cores of 3.0GHz quad-core Intel Xeon E5450 processors with 2Gb RAM per core
- gen1 is 128 cores of 1.6GhZ AMD Opteron 242 processors with 1Gb RAM per core
Details of the software available on Cheaha can be found on the Cheaha cluster configuration page, an overview follows.
Cheaha's software stack is built with the ROCKS, a Linux-based cluster distribution. Cheaha's operating system is CentOS with the following major cluster components:
A summary of the available computational software and tools available includes:
- Intel Compilers
- GNU Compilers
Cheaha is connected to the UAB Research Network which provides a dedicated 10Gbs networking backplane between clusters located the UAB Shared Computing Facility and Department of Computer and Information Science HPC Center. At present only Cheaha and Ferrum are connected via these 10Gbs interfaces. Data transfers rates of almost 8Gbs between these hosts have been demonstrated using Grid FTP, a multi-channel file transfer service that is used by GridWay to move data between clusters as part of the job management operations. This performance promises very efficient job management and the seamless integration of other clusters as connectivity to the research network is expanded.
The continuous resource improvement process involves collecting benchmarks of the system. One of the measures of greatest interest to users of the system are benchmarks of specific application codes. The following benchmarks have been performed on the system and will be further expanded as additional benchmarks are performed.
Cheaha uses Ganglia to report cluster performance data. This information provides a helpful overview of the current and historical operating stats for Cheaha. You can access the Ganglia monitoring page here.
Cheaha is a general-purpose computer resource made available to the UAB community by UAB IT. As such, it is available for legitimate research and educational needs and is governed by UAB's Acceptable Use Policy (AUP) for computer resources.
Many software packages commonly used across UAB are available via Cheaha. For more information and introductory help on using this resource please visit the resource details page.
To request access to Cheaha, please send a request to send a request to the cluster support group.
Cheaha's intended use implies broad access to the community, however, no guarantees are made that specific computational resources will be available to all users. Availability guarantees can only be made for reserved resources.
Secure Shell Access
Please configure you client secure shell software to use the official host name to access Cheaha:
To ensure that you are connecting to the legitimate host, you can verify that the fingerprint presented by your secure shell client for cheaha.uabgrid.uab.edu matches:
Upgrade Note: The previous host name (cheaha.ac.uab.edu) is mapped to the new hostname to support the transition for existing users. If you connect using the old host name, this name mapping may trigger a warning in your secure shell software about the host fingerprint change. The preceding finger print enables you to confirm you are indeed connecting to the legitimate Cheaha interface.
The storage systems from the original Cheaha (cheaha.ac.uab.edu) system are scheduled for decommission on March 1 2009. After this date they will no longer be available.
This change affects the home directories of all users who had accounts on the original Cheaha system (cheaha.ac.uab.edu). Affected users should copy any data they want to preserve from their old home directories to their current home directory on Cheaha by March 1 2009.
As a convenience, the old home directories are accessible directly via the file system of the new system. The old files are available as read-only data to ensure that new data is not saved to these locations. Users should copy any files they wish to preserve from their old home directories located at /oldcheaha/$LOGNAME to their current home directory (located at $HOME).
The following command will copy all files to a folder call "oldcheaha" under your current home directory. Users are encouraged, however, to use this opportunity to selectively copy only the data they wish to preserve.
cp -a /oldcheaha/$LOGNAME $HOME/oldcheaha
Cheaha provides performance and management improvements for scientific workflows by enabling access to the processing power of multiple clusters through the use of the GridWay scheduling framework. GridWay enables the development of scientific workflows that leverage all computing resources available to the researcher and that can be controlled through a single management interface. This feature puts state-of-the-art technology in the hands of the research community.
Enhancements to Cheaha in general, and, in particular, its scheduling framework are intended to remain transparent to the user community. Cheaha is first and foremost a resource for predictable and dependable computation. In this spirit, Cheaha can continue to be viewed as a traditional HPC cluster that supports job management via Sun Grid Engine (SGE). User's who have no need for or interest in maximized access to computational resources can continue using the familiar SGE scheduling framework to manage compute jobs on Cheaha, with the familiar restriction that SGE managed jobs can only leverage the processing power of Cheaha's local compute pool. That is, these jobs, as in the past, cannot leverage cycles available on other clusters.
Cheaha provides access to its local compute pool via the SGE scheduler. This arrangement is identical to the existing HPC clusters on campus and mirrors the long-established configuration of Cheaha. Researchers experienced with other SGE-based clusters should find no difficulty leveraging this feature. For more information on getting started with SGE on Cheaha please see the cluster resources page.
Cheaha provides enhanced scientific workflow management and development capabilities via the GridWay scheduling framework. GridWay enables the orchestration of scientific workflows across multiple clusters. The pool of resources available as part in Phase 1 includes Cheaha, Olympus, Everest, and Ferrum. This pool can be monitored by any user of Cheaha by executing the gwhost command when logged into Cheaha.
The GridWay framework provides two interfaces. A scheduler interface similar to SGE is recommended for initial exploration and ordinary use. The scheduler activity can be monitored with the `gwps` command. Job submit and monitoring commands initiate and control commands described in a job description file. Outside of slightly different commands, the job description file operates as a template where specific fields are populated to affect the operation of the scheduler. These templates are less ambiguous than traditional SGE job scripts and can provide a direct migration path from SGE. A more subtle difference is that an explicit (though automated) job staging step is involved in order to start jobs. This can require more explicit handling of input and output files than is ordinarily required by SGE.
Additionally, very powerful programatic control is available via the DRMAA API. DRMMA enables the development of advanced scientific workflows that can leverage any number of computational resources. GridWay provides bindings to many popular programming languages like C, Java, Perl, Python and Ruby.
Both the traditional scheduler-based interface and the DRMMA API have been explored by development groups on-campus during the pilot evaluation phase of GridWay.
- UAB IT, UAB's School of Public Health Section on Statistical Genetics (SSG), and CIS established an interest group for R which leveraged the scheduler-based interface to improve the performance of select R language statistical analysis workflows.
- The CIS Collaborative Computing Laboratory has heavily leveraged the Java DRMMA API to develop DynamicBLAST, a scientific workflow to orchestrate and maximize the performance of BLAST across multiple resources.
Adoption of GridWay is encouraged and future compute capacity enhancements will leverage the inherent flexibilities of GridWay. The nature of any new technology, however, implies a learning curve. The learning curve need not be steep and direct migration of basic SGE scripts is possible. Additionally, all Cheaha accounts are configured to support the use of GridWay as an alternative scheduler, empowering the adventurous.
Some important points worth considering in evaluating adoption of GridWay.
- GridWay cannot perform magic. If you ordinarily do not have access to other clusters or your code does not (or will not) run on a targeted cluster, GridWay cannot solve these problems for you. You must ensure your codes run on all compute resources you intend to include in your scheduling pool prior to submitting jobs to those resources.
- Migration of MPI jobs to a multi-cluster environment may involve additional effort. If you simply use MPI to coordinate the workers (rather than for low-latency peer communication), you should generally be able to structure your job to work across cluster boundaries. Otherwise, additional effort may be required to divide your data into smaller work units. It should be noted, however, that MPI itself cannot be used to communicate across cluster boundaries. Jobs distributed across cluster boundaries that leverage MPI internally must be sub-divided to run within isolated communication domains.
The UABgrid User Community stands ready to help you with GridWay adoption. If you are interested in exploring or adapting your workflows to use GridWay please subscribe to the UABgrid-User list and ask all questions there. Please do not submit GridWay-related questions via the ordinary cluster support channels. Additional on-line documentation will be developed to provide migration examples to help you further explore the power of GridWay.
Operational support for Cheaha is provided by the High Performance Computing Support group in UAB IT. For questions regarding Cheaha or to initiate support requests, please send your request to firstname.lastname@example.org.