Skip to content

Files

Latest commit

 

History

History
70 lines (61 loc) · 3.73 KB

SPECIFICATION.md

File metadata and controls

70 lines (61 loc) · 3.73 KB

Specification

The specifications below quantify Firecracker's promise to enable minimal-overhead execution of container and serverless workloads. These specifications are enforced by integration tests (that run for each PR and main branch merge).

On an M5D.metal instance (with hyperthreading disabled) and an M6G.metal instance and given host system resources are available (e.g., there are enough free CPU cycles, there is enough RAM, etc.), customers can rely on the following:

  1. Stability: The Firecracker virtual machine manager starts (up to API socket availability) within 8 CPU ms1 and never crashes/halts/terminates for internal reasons once started. Note: The wall-clock time has a large standard deviation, spanning 6 ms to 60 ms, with typical durations around 12 ms.
  2. Failure Information: When failures occur due to external circumstances, they are logged2 by the Firecracker process.
  3. API Stability: The API socket is always available and the API conforms to the in-tree Open API specification. API failures are logged in the Firecracker log.
  4. Overhead: For a Firecracker virtual machine manager running a microVM with 1 CPUs and 128 MiB of RAM, and a guest OS with the Firecracker-tuned kernel:
    • Firecracker's virtual machine manager threads have a memory overhead <= 5 MiB. The memory overhead is dependent on the workload (e.g. a workload with multiple vsock connections might generate a memory overhead > 5MiB) and on the VMM configuration (the overhead does not include the memory used by the MMDS data store.

      The overhead is tested as part of the Firecracker CI using a memory cop.

    • It takes <= 125 ms to go from receiving the Firecracker InstanceStart API call to the start of the Linux guest user-space /sbin/init process. The boot time is measured using a VM with the serial console disabled and a minimal kernel and root file system. For more details check the boot time integration tests.

    • The compute-only guest CPU performance is > 95% of the equivalent bare-metal performance. [integration test pending]

  5. IO Performance: With a host CPU core dedicated to the Firecracker device emulation thread,
    • the guest achieves up to 14.5 Gbps network throughput by using <= 80% of the host CPU core for emulation. [integration test pending]
    • the guest achieves up to 25 Gbps network throughput by using 100% of the host CPU core for emulation. [integration test pending]
    • the virtualization layer adds on average 0.06ms of latency. [integration test pending] See further details on network performance
    • the guest achieves up to 1 GiB/s storage throughput by using <= 70% of the host CPU core for emulation. [integration test pending]
  6. Telemetry: Firecracker emits logs and metrics to the named pipes passed to the logging API. Any logs and metrics emitted while their respective pipes are full will be lost. Any such events will be signaled through the lost-logs and lost-metrics counters.

Footnotes

  1. CPU ms are actual ms of a user space thread's on-CPU runtime; useful for getting consistent measurements for some performance metrics.

  2. No logs are currently produced in the span of time between the jailer process start-up and the logging system initialization in the firecracker process.