Thanks John and Kaz for submitting this. Let me know if you need any help on Wally. I've updated the Teradimms to the latest IBM released firmware: 1.5.0.35, and kernel module: 1.5.0.41. Ben
On Mar 18, 2015, at 5:59 PM, John Tramm <[email protected]> wrote:
From: John Tramm <[email protected]> Exploring the Viability of NVRAM for use in Task-Based 3D Method of Characteristics (MOC) Style Nuclear Reactor Simulation Applications. John Tramm, Kazutomo Yoshii MCS Many new upcoming HPC architectures feature large NVRAM storage on-node, offering a huge boost to memory capacity provided that the application's algorithms can handle the high latency involved in NVRAM access while also accommodating the low durability of NVRAM DIMMs.
In light of the potential benefits that NVRAM provides, we are investigating the usage of high capacity NVRAM modules for use in task-based 3D Method of Characteristics (MOC) full-core nuclear reactor simulation applications, such as SimpleMOC and OpenMOC. NVRAM offers large potential benefits to this style of application as each node requires a large data footprint, yet this data only needs to be accessed once per computational sweep. The amount of computation required between each read/write from NVRAM is predicted to be extremely high. This fact, combined with the task-based work model of SimpleMOC, means that masking NVRAM latency is feasible with only minor algorithmic improvements, and also that durability requirements can be met.
Additionally, the large data footprint of 3D MOC applications is also required to be exchanged between node neighbors in a ghost cell exchange pattern following each computational sweep. In the case of SimpleMOC, the ghost cells compose nearly the entire data footprint of the application, which allows for vastly fewer nodes to be required by the computation if on node NVRAM modules can be utilized effectively. While ghost cell exchanges may be difficult to directly execute on a single node or few node testbed system, there are still useful metrics we can gather in order to model the altered costs of ghost cell exchanges given the use of NVRAM over traditional DRAM.
The primary goal of this study is to implement NVRAM usage techniques into the SimpleMOC application in order that we may quantify the costs/rewards of NVRAM usage on 3D MOC reaction simulation algorithms. With these latency and performance parameters, we will be able to build a performance model for the application when run on future many-node HPC systems. This study is done to better prepare nuclear simulation applications for usage on exascale class supercomputers as part of the Center for Exascale Simulation of Advanced Reactors (CESAR) project. CESAR (MCS) MIT John Tramm, Kazutomo Yoshii Wally (for NVRAM exFlash DIMMs) [Schedule (when access is needed and for how long):]
-- This e-mail was sent from a contact form on The Joint Laboratory for System Evaluation (http://press3.mcs.anl.gov/jlse)
_______________________________________________ Admins mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/admins