1. System Overview

1.1. System Configutation

The Quantum Simulator System consists of compute nodes for simulation (consisting of 512 FX 700), a jump server for login, a login server for job submission, a shared directory, a job scheduler, a private PyPI repository, a document server, and a proxy server for Internet access.

The user logs in to the login server through the jump server. The compute nodes are available from the login server through the job scheduler Slurm. The compute nodes are connected by InfiniBand, and the quantum circuit is simulated while performing parallel calculation using MPI (Message Passing Interface).

The information for each server/node that makes up the quantum simulator system is as follows.

The role and IP address/hostname of each server

Name	IP / Host	Explanation
Jump Server	106.184.62.10 Port number: Notification from Fujitsu	Server in a DMZ
Login Server	login-server	User environment for utilizing computation Accessible from the jump server
Compute Node	N/A	Cluster system for executing quantum program
Shared Directory	N/A	Directory for placing files shared between compute nodes
Private PyPI Repository	simulator-pypi	Private Python package repository for Fujitsu software packages (e.g. mpiQulacs)
Document Server	simulator-docs	Server that contains updated information and API manuals
HTTP Proxy Server	simulator-proxy	Proxy server used to access Internet from the compute node

Specifications for each server/node

Server name	Item
Login Server	OS	Rocky Linux 8.6
	CPU	Intel(R) Xeon(R) Gold 6348 (2.6GHz) x 1core
	Memory	32GB
Compute Node	OS	Rocky Linux 8.5
	CPU	Fujitsu A64FX (2.0GHz) x 48cores
	Memory	32GB
Shared Directory	File System	GlusterFS
	Storage Capacity	3TB/group

Note

Use the jump server only as a relay to access the login server. Do not create or execute programs on the jump server.

Note

The initial storage capacity of the shared directory is 3 TB, but if you still run out of space after deleting unnecessary files, contact your Fujitsu representative.

1.2. Job Scheduler

The job scheduler is a system that divides multiple shared computer resources so that multiple users can execute jobs and manages them so that each user can occupy the compute node for a certain period of time. A job is a set of shell processing units that, when a command or script is submitted as a job, is executed when the requested resources are ready.

This system uses Slurm Workload Manger (Slurm) as the job scheduler.

The following figure shows the flow from the login server to the calculation node via the job scheduler Slurm, and the connection to the shared directory.

① Log into the login server (loginvm- < three digits above group ID >)

② Run the Slurm command on the login server

③ Job scheduler schedules jobs based on the command and number of nodes required

④ Job scheduler submits jobs to compute nodes

⑤ Job runs on compute nodes

Note

The home directory of each user’s login server mounts the shared directory. The same directory is mounted on all compute nodes.

The following is an image of ③ job scheduler scheduling.

The job scheduler uses queues to manage jobs submitted by users.

Interactive queue is a queue for logging in to compute nodes and executing commands interactively. It is mainly used to create an MPI execution environment on compute nodes, compile scripts, and check the normal operation before batch execution. It is assumed to allocate one or two nodes at a time in the Interactive queue.

Batch queue is a queue for submitting scripts from the login server to the compute nodes.

We will inform you separately about the total number of nodes available for Interactive and Batch queues.

Note

The number of nodes used and the maximum number of qubits are as follows, and quantum circuits up to 39 qubits can be executed when 512 nodes are used.

Number of nodes used	512	256	128	64	32	16	8	4	2	1
Maximum qubit number	39	38	37	36	35	34	33	32	31	30