"K Pre-Post Cloud" is a private cloud as an experimental platform in the supercomputing environment in order to enhance pre/post data processing features including data analysis and visualization. (The service was launched in FY2018). In Supercomputer Fugaku as well as the K computer, we continue to provide the experimental service to obtain technical knowledge and know-how of operation for pre/post servers installed in the supercomputer environment.
K Pre-Post Cloud is recommended for various users who want to
A user can choose virtualized resources (vCPUs, RAM, storage, and so on) based on template-style "flavors" that define the resource configuration.
To easily control a set of the resource sizes, we provide "flavors" as a definition of assigned resources based on the table as follows. For example, "A5.medium" means that the VM consists of 24vCPUs, 64GiB, and 512GiB SSD. In the table, A1-A8 (They define vCPUs-RAM size.) are standard instance types. B1-B5 are memory-oriented instance types. C1-C6 are compute-oriented instance types. Also, we provide new flavors ("A8.huge.gpu-p" and "A8.huge.gpu-v") that are special flavors to use a GPGPU instance.
The flavor's postfix (tiny, small, medium, large, and huge) means "Root Disk" storage size. The "Root Disk" is based on RAID 0 volume with SSD to provide fast I/O and is designed for ephemeral space (i.e. operating system installation space and temporary workspace). When a user shuts down his/her VM, the VM's Root Disk is deleted at the same time. (It's so-called ephemeral.)
Note: The following contents includes obsolete information regarding the K computer. We will update them in the future.
In the K Pre-Post Cloud service, there are several types of storage available. As local storage on the virtualized environment, a SSD (root disk) is suitable for running instance. The Ceph storage as a secondary storage is designed for storing VM images/snapshots while its space is not enough in size to keep all program outputs. If a user needs to use larger space to store his/her simulation results, we recommend using GFS on K. Also, the current quotas as default is tentative. If you need to change the quotas, please contact us.
Ansible playbook to use GFS space is provided through the local GitLab site. Please refer the site via a VPN connection.
In the hardware aspect, the private cloud uses 25GbE network switches to connect among VMs and the Ceph storage space. In addtion, user's VM can access the GFS space and the Internet via several physical gateway servers with redundant 10GbE interfaces. Meanwhile, in the software aspect, a user can assign floating IP addresses to his/her VMs in order to access them from outside of the private cloud via a software-defined router. Note that the floating IP address is a private IP address. Thus, as long as the VPN connection establishes, the user's VMs with floating IP addresses can be accessed from the Internet. (We don't provide global IP addresses due to the RIKEN security policy.)
In order to give more users an opportunity to use the private cloud, we introduce a simple mechanism that automatically terminates old VMs in a given period of time depending on the flavors. This policy is based on that a bigger resource consumer tends to be imposed short duration. Meanwhile, the policy allows smaller VMs to live longer. In the table, "inf" means that VMs with the flavors are allowed to survive until the end of the fiscal year or the expiration date. Why we introduce the policy is, whereas commercial cloud services collect a fee based on the amount of the actual usage, K Pre-Post Cloud does not charge anything for the service. Thus, we need to introduce the mechanism to provides liquidity for VMs. (1w: 1week, 2w: 2weeks, 4w: 4weeks, inf: the end of the fiscal year or the expiration date)
|K Pre-Post Cloud|
|CPU||Intel Xeon Platinum 8168 (Skylake) (24cores/2.7Ghz/33MB) x 2 (/node)|
|#nodes||11 (compute nodes)|
|Total #cores||528 cores
(1056 vCPUs, Hyper-Threading enabled)
|Storage||SSD(9.6TB/node) + Ceph(150TB) + others (peripheral storage in K and Fugaku)|
|GPGPU||Part of the compute nodes has GPGPUs.
NVIDIA Tesla P100 (16GiB) x 4 / system
NVIDIA Tesla V100 (16GiB) x 4 / system
|OS||HostOS: RHEL 7.4
GuestOS: CentOS7, Ubuntu16.04LTS/18.04LTS
(A user can choose a guest OS from the image list.)
|misc||A service portal provides various interfaces (Web/CLI/REST API) to control his/her VM. Through the interface, a user can get his/her VM on demand.|
This material explains how to create a GPGPU instance.
If you have any questions about getting started, please feel free to ask.