From c492528e6e77aac3b4a1bc5dedcd8024189ad81c Mon Sep 17 00:00:00 2001 From: HGY <42902781+hugy718@users.noreply.github.com> Date: Mon, 29 Jul 2024 11:17:15 +0800 Subject: [PATCH] Update docs (#72) --- .gitmodules | 6 ------ docs/source/_static/task_executor.md | 27 +++++++++++++++++++++++++-- 2 files changed, 25 insertions(+), 8 deletions(-) delete mode 100644 .gitmodules diff --git a/.gitmodules b/.gitmodules deleted file mode 100644 index 8be367b..0000000 --- a/.gitmodules +++ /dev/null @@ -1,6 +0,0 @@ -[submodule "sdk/example_containers/sgx-task/deps/llhttp"] - path = sdk/example_containers/sgx-task/deps/llhttp - url = https://github.com/nodejs/llhttp.git -[submodule "sdk/example_containers/sgx-task/deps/libuv"] - path = sdk/example_containers/sgx-task/deps/libuv - url = https://github.com/libuv/libuv.git diff --git a/docs/source/_static/task_executor.md b/docs/source/_static/task_executor.md index 5fe4db3..ce4fd61 100644 --- a/docs/source/_static/task_executor.md +++ b/docs/source/_static/task_executor.md @@ -10,11 +10,22 @@ A task request sent over from MECA contains the id that defines the computation The task executor can manage multiple tasks with different resource usage intents at a given time. Therefore, internally the task executor manages the tasks with an id constructed from both the id and resource in a MECA request. -By default, if the request does not specify the resource usage intent, it is assumed to work with the default resource usage: 1 CPU core and 128MB RAM. The specified CPU should be a float with one decimal and the ram RAM is an integer. +By default, if the request does not specify the resource usage intent, it is assumed to work with the default resource usage: 1 CPU core and 128MB RAM. The specified CPU should be a float with one decimal and the RAM is an integer in the unit of MB. + +## Isolation support + +Task executor now by default uses Docker for isolation. It can also use MicroVM and SGX with additional configurations. The servers to run task with MicroVM or SGX should have the necessary platform software to do so. + +* MicroVM: support by [Kata-container][1]. The server shall install kata-container and register kata container runtime to Docker [2][2]. +* SGX: support by [Linux-SGX][3]. The server shall enable SGX support via BIOS and install all the Intel platform software packages, And ensure the Intel AESM service running with systemd. + +## GPU support + +Nvidia GPU support uses the `NVIDIA/go-nvml` package to provision local detect and manage GPU devices. The GPU devices shall be mounted to the task executor container so that it is schedulable [4][4]. ## Cleaning -A task is maintained as a running server on the host machine. It is periodically cleaned by the task executor if they are inactive for a configured threshold of time (default 1min). +A task is maintained as a running server on the host machine. It is periodically cleaned by the task executor if they are inactive for a configured threshold of time (default 1min). When there is resource shortage to schedule a new incoming task, previous tasks kept running but not handling request will be removed to free up resources. ## Configuration @@ -33,3 +44,15 @@ curl http://172.18.0.255:2591/pause -X POST curl http://172.18.0.255:2591/update-config -X POST -H "Content-Type: application/json" -d '{"timeout": 2, "cpu":2, "mem":4096, "microVM_runtime":"kata"}' curl http://172.18.0.255:2591/unpause -X POST ``` + +## References + +1. [KataContainers][1] +2. [Docker support for Kata Container][2] +3. [Intel SGX for Linux][3] +4. [Docker GPU support][4] + +[1]: https://katacontainers.io/ +[2]: https://docs.docker.com/engine/alternative-runtimes/ +[3]: https://github.com/intel/linux-sgx +[4]: https://docs.docker.com/config/containers/resource_constraints/#gpu