In our previous newsletter, we described how some of our HPC clusters in the shared resource pool are now running with compute nodes from the cloud. This was made possible with the implementation of a virtual private network between NUS and AWS, our cloud solution provider, extending our HPC into the cloud in a seamless manner. Job scheduling is still managed by PBS Pro, and the usual queuing is still necessary to ensure that jobs are scheduled in the cloud compute nodes in a fair manner.

For those who do not want to wait, we are now making available the paid subscription of the compute nodes in the cloud. With this new subscription, HPC users can use their own funds to spin up compute instances suitable for their own use, without having to submit to a job queue. Those interested to use this pilot subscription service may contact us via nTouch, and we will have a more detailed conversation on the on-boarding process.

HPC users may be familiar with our on-premise HPC condominium service. The cloud account will be a replacement of the HPC condominium service, and all subscribers will need to register for an account with us. This subscription is open to all NUS staff and will require the endorsement of their individual department finance administrator to ensure that proper invoicing can be arranged so that payments can be made directly to AWS. This is an improvement to the usual credit card method that AWS operates for general cloud users, because now, the department finance administrators can have proper paperwork with invoice number attached to the monthly expenditures of the researcher’s account.

The compute resources in the cloud come in the form of compute instances (similar to virtual machines) and for HPC users, most common instance types would be the compute-optimised and the memory-optimised instances. In general, the cloud provides other general purpose instance types, which may be more suitable for less compute intensive workloads such as web servers, but these may not be the top choice for HPC users. Instance types also have different configurations ranging from 2 vCPUs, up to 72 vCPUs within a single compute-optimised instance and up to 96 vCPUs within a single memory-optimised instance. Compute-optimised instances come with 2GB of RAM per vCPU, while memory-optimised instances have higher RAM, at 8GB per vCPU. Table 1 below shows a sample of currently available compute-optimised and memory-optimised instances:

Table 1

As seen, the amount of compute resources is quite substantial, and researchers can run their HPC workloads quite effectively in the cloud.

The cloud services come in several pricing models. The basic pricing is the on-demand pricing, which allows researchers the flexibility to spin up instances in the cloud as and when they need them. This flexibility comes at a premium and is the costliest pricing option in the cloud. To mitigate the high cost, one should not leave the resources running idle, and should be shutdown whenever there is no workload running. On-demand instances would be suitable for ad-hoc or one-off computational jobs that may be needed urgently.

If a compute instance is going to be running workloads for a longer period of time, such as one year or three years, the cloud service provider has an option for reserved instances which can be acquired at a lower price ranging from 50% to 61% lower than on-demand pricing.

For special types of workloads that are short and interruptible, the cloud service provider offers the “spot” instance pricing, where researchers can bid for idle spare resources in the cloud at a price that they are willing to pay. Once an instance is available at, or below, the price that is bid by the subscriber, the instance will be made available for the subscriber to run. Table 2 below illustrates the difference in the costs based on the various pricing schemes described above:

Table 2

Researchers will be given full access to the instances that are provisioned in the cloud, and they will be able to install any software in the system. Therefore, installation of general, open-sourced software will be rather straight-forward. For licensed software, please contact us at via nTouch, as there may be some network configurations that may need to be set up before networked licenses can be granted to the cloud instances.

If you are interested to run your HPC workloads in the cloud, send your requests to us via nTouch, with details of your compute requirements. We will evaluate your computational needs and your budget to come up with a suitable solution for your HPC workloads.

nTouch (for desktop web browser clients):

Or navigate via the “Browse categories” tab: Business Requests > High Performance Computing > HPC Enquiries.