Instance Flavors¶

Jetstream2 provides a range of instance flavors (also known as sizes), each under three different types with different capabilities. These instances consume an allocation’s service units (SUs) according to how much physical hardware they occupy. Larger instance flavors consume SUs at a greater rate.

The unit of allocation for Jetstream is based on a virtual CPU (vCPU) hour: 1 service unit (SU) is equivalent to 1 vCPU for 1 hour of wall clock time. The tables below outline the instance types and flavors for Jetstream2.

While the root disk sizes are fixed for each instance flavor, there is an option called “volume-backed” that allows you to specify a larger root disk, using quota from your storage allocation. Instructions for this are in the user interface-specific documentation for creating an instance (Exosphere, Horizon, CLI).

Jetstream2 CPU¶

Flavor	vCPUs	RAM (GB)	Local Storage (GB)	Cost per hour (SU)
m3.tiny	1	3	20	1
m3.small	2	6	20	2
m3.quad	4	15	20	4
m3.medium	8	30	60	8
m3.large	16	60	60	16
m3.xl	32	125	60	32
m3.2xl	64	250	60	64
m3.3xl*	128	500	60	128

* m3.3xl are not available by default. This flavor is available by request and with proper justification.

Jetstream2 Large Memory¶

Jetstream2 Large Memory instances have double the memory (RAM) of equivalently-resourced CPU instances. They cost 2 SUs per vCPU hour, or 2 SUs per core per hour.

Flavor	vCPUs	RAM (GB)	Local Storage (GB)	Cost per hour (SU)
r3.large	64	500	60	128
r3.xl	128	1000	60	256

Jetstream2 GPU¶

Jetstream2 GPU instances include a partial or full NVIDIA A100 GPU, with up to 40 GB of GPU RAM. Jetstream2 GPU instances cost 4 SUs per vCPU hour, or 4 SUs per core per hour.

Flavor	vCPUs	RAM(GB)	Local Storage (GB)	GPU Compute	GPU RAM (GB)	Cost per hour (SU)
g3.small	4	15	60	20% of GPU	5	16
g3.medium	8	30	60	25% of GPU	10	32
g3.large	16	60	60	50% of GPU	20	64
g3.xl	32	125	60	100% of GPU	40	128

Note: If you are using a partial-GPU flavor, and the remainder of the underlying physical GPU is idle, your instance may provide higher compute performance than the flavor strictly allots. In other words, the GPU compute for each flavor is a minimum value, while the GPU RAM is a maximum value.

Jetstream2 instance types, flavors, and associated policies are subject to change in the future.

---

Example of SU estimation:¶

Note: You can now estimate your SU needs using the usage estimation calculator here: Usage Estimation Calculator

• First determine the compute resource appropriate to your needs (CPU only, large memory, GPU):
  • If your work requires 24 GB of RAM and 60 GB of local storage:
    • you would request 8 SUs per hour to cover a single m3.medium instance.
  • If your work requires 10 GB of local storage in 1 core using 3 GB of RAM:
    • you would request 2 SUs per hour for an m3.small instance.
  • If your work requires 1TB of RAM:
    • you would request 256 SUs per hour for an r3.xl instance on Jetstream Large Memory
  • If you work requires 20 GB of GPU RAM:
    • you would request 64 SUs per hour for a g3.large instance on Jetstream GPU
• You then would calculate for the appropriate resource (refer to the tables above):
  • For Jetstream2 CPU, you would then multiply by the number of hours you will use that flavor instance in the next year and multiply by the number of instances you will need.
  • For Jetstream2 Large Memory and GPU, either refer to the SU cost per hour in the last column, or multiply hours times 2 for LM or 4 for GPU
• To calculate the number of SUs you will need in the next year, first estimate the number of hours you expect to work on a particular project. For example, if you typically work 40 hours per week and expect to spend 25% of your time on this project that would be 10 hours per week.
• Next, calculate the total number of hours per year for this project:
  • Total hours = 10 hours per week * 52 weeks per year
  • Total hours = 520
• Finally, calculate the total SUs for the year for a single instance:
  • Total SUs = 520 hours per year * vCPUs
    • e.g. For a Medium instance: Total SUs = 520 hours per year * 8vCPUs
    • Total SUs = 4160
• If your project requires more than 1 instance, multiply the total SUs by the number of instances that you will need:
  • Total SUs needed for 3 medium flavor instances = 3 * 4160
  • Total SUs = 12480

The calculations above assume that your instance is shelved when not in use.  For instructions see:

• Cacao instance management actions
• Exosphere instance management actions
• Horizon instance management actions
• Command line instance management actions

SU Estimation for Infrastructure or “Always On” allocations¶

For jobs that may need to run for extended periods or as “always on” infrastructure, you can take this approach:

instance cost (SUs) x 24 hours/day x 365 days = single instance cost per year

or as an example for each resource, an m3.large, r3.large, and g3.large each running for a year:

    m3.large (16 cores) x 24 hours/day x 365 days = 140,160 SUs
    r3.large (64 cores x 2 SUs/hour) x 24 hours/day x 365 days = 1,121,280 SUs
    g3.large (16 cores x 4 SUs/hour) x 24 hours/day x 365 days = 560,640 SUs