2022 ERCAP Guide

NOTE: This page highlights the timelines and considerations especially for ERCAP 2022 submissions. For an overview of the general process, eligibility, a list of allocation managers, and other guidance please see the Allocations Overview page.

ERCAP Timeline

• Open: September 7, 2021
• Close: October 4, 2021
• ERCAP Office Hours:
• Thursday, September 9
• Thursday, September 30
• Monday, October 4 (ERCAP due date)
• Office hours will take place from 9-12 noon and 1-4 pm (Pacific time). You can get the Zoom access information here. (NOTE, you will be required to log in with your NERSC username, password, and MFA in order to view this information.)

Changes for 2022

Separate Allocations for GPU-Accelerated and CPU-Only Resources

There will be two separate compute allocation pools: one for the Perlmutter GPU-accelerated nodes and another for all CPU only nodes (Perlmutter CPU-only nodes and all Cori nodes, both Haswell and KNL).

Allocation Requests: Node Hours

Allocations, charging and usage tracking will be in units of node hours in 2022. There are two distinct types of node hours: "CPU Node Hours" and "GPU Node Hours." The basis for one CPU node hour is time spent running on one Perlmutter CPU-only node for one hour. Time charged for running on Cori KNL and Haswell nodes will be scaled appropriately to reflect the relative performance of those nodes on a "typical" code run at NERSC. 

Cori and Perlmutter CPU-Only Node Hours

In 2021 NERSC used the unit of "NERSC Hour" for allocations and charging. One node hour of computing on Cori KNL is valued at 80 NERSC Hours and one hour of computing on one Cori Haswell node is valued at 140 NERSC Hours. Use the table below how to calculate the value of a node hour of computing in the 2021 and new 2022 units. Note these are base charges and that actual charges can be affected by running jobs in certain discount or premium queues.

Example

Let's say that in 2021, you used 1 million NERSC-hours on Cori. If you plan to use approximately the same amount of (non-GPU) compute resources in 2022, you would want to request

1,000,000 ÷ 400 = 2,500 node-hours.

Perlmutter GPU Node Hours

A GPU Node-hour is defined as occupying a single Perlmutter GPU node (consisting of a single socket of an AMD EPYC 7763 (Milan) processor and four NVIDIA Ampere A100 GPUs: link) for one hour. A job that runs across 500 nodes for 1.5 hours therefore uses 500 x 1.5 = 750 GPU node-hours.

Perlmutter Timeline

NERSC has not yet announced when we will begin charging for usage of Perlmutter. We're still asking for your best estimates of what you will use on Perlmutter in 2022. Don't worry about making a perfect estimate. 

Codes Provided on Perlmutter

We plan to provide some common libraries and applications on Perlmutter. Please see the Applications and Libraries documentation pages for more details.

GPU Readiness

We’re asking you to assess the GPU readiness of the codes you plan to use on the Perlmutter GPU nodes. There are a wide range of discussions that are possible to describe GPU readiness for our codes and user base. So, please provide the relevant information about your current GPU status, plans, goals and/or limitations to help us understand the current state of your codes and what will best serve your success in the future.

 As part of your discussion, please answer any questions in this list that are relevant to your GPU readiness, including details that will help us better understand your GPU readiness status:

• Is your software, workflow, or code development ready to be primarily focused on GPUs?
• At this time, would CPU or GPU systems (i.e., Cori vs. Perlmutter GPU) better serve your scientific goals (for example, in terms of success, funding potential, porting commitment)?
• At this time, would CPU or GPU systems more efficiently serve your scientific goals (for example, in terms of time to solution or the above given charge rates)?
• If you don't have a GPU-ready code at this time, are you planning to develop your code under this ERCAP request, or do you plan to run a GPU-ready code that is currently in development by another code group? What is your expected timeline for GPU readiness?
• What is currently the biggest hindrance to your productivity on GPUs (e.g., training, software bug, hardware limitation, time to work)?

For additional information on preparing to run on GPUs, visit our Perlmutter Readiness page.

Questions/Concerns?

Submit a ticket or make an appointment.