Polyphemus

A server for managing FPGA build and execution environments.

Execution Model

Polyphemus is built to run on heterogenous environments where several machines with different capabilities co-operate to complete jobs.In our use case, we use several AWS [FPGA AMI][fpga-ami]s to synthesize hardware and then run those on a single F1 instance.

In development, Polyphemus has two parts: a server and several WorkProcs. The server is responsible for receiving jobs, adding them to the database, and generate the UI. Each WorkProc runs several threads with a specific 'stage'. Each stage is responsible for moving finishing a part of a job based on its capability.

For example, an AWS deployment might have 4 synthesis server each of which can run on synthesis job while there is one F1 server for executing the jobs.

Running Polyphemus

Prerequisites

• Install pipenv
• Run pipenv install to configure the environment.

Configuration

The server keeps the data, including the database and the file trees for each job, in an instance directory here. To configure Polyphemus, create a file called polyphemus.cfg in the instance directory. For an exhaustive list of options, see their default values.

These ones are particularly important:

• TOOLCHAIN: Polyphemus supports two Xilinx HLS workflows: SDAccel (on Amazon F1) and SDSoC. Set this to "f1" for deployment on F1. Set it to anything else to use the SDSoC workflow.
• PARALLELISM_MAKE: The number of jobs to process in parallel in the "make" stage. The default is 1 (no parallelism).
• HLS_COMMAND_PREFIX: A prefix to use for every command that requires invoking an HLS tool. Use this if you need to set up the environment before calling make, for example. This should be a list of strings.

Development

Run this command to get a development server, with Flask's debug mode enabled:

$ FLASK_APP=polyphemus.server FLASK_ENV=development pipenv run flask run

You can also use make dev as a shortcut.

This route automatically starts the necessary worker threads in the same process as the development server.

Deployment

For proper production, there are two differences from running the development version: You'll want to use a proper web server, and Polyphemus will want to spawn a separate process just for the worker threads.

Use this command to start the workers:

$ pipenv run worker

For the server, Gunicorn is a good choice (and included in the dependencies). Here's how you might invoke it:

$ pipenv run gunicorn polyphemus.server:app

The make serve target does that.

The two processes communicate through a Unix domain socket in the instance directory. You can provide a custom instance directory path to the workproc invocation as an argument.

Multi-machine Deployment

When deploying on multiple machines, we'll start a single instance of the server and start several workers, each with machine-specific capabilites, on every machine.

TODO: Finish this section after deployment testing on F1.

Using Polyphemus

There is a browser interface that lets you view jobs and start new ones. It's also possible to do everything from the command line using curl.

Submitting Jobs

To submit a job, upload a file to the /jobs endpoint:

$ curl -F [email protected] $POLYPHEMUS/jobs

For example, you can zip up a directory and submit it like this:

$ zip -r - . | curl -F file='@-;filename=code.zip' $POLYPHEMUS/jobs

If the directory contains data files with .data extension, they'll be copied over to the target FPGA.

Job Options

When submitting a job, you can specify job configuration options as further POST parameters. Some options are only relevant for a particular HLS workflow (see "Configuration," above). The options are:

• For all workflows:
  • skipexec, to avoid actually trying to run the generated program. (Only necessary when estimate is false---estimated runs skip execution by default.)
  • make, to use a Makefile instead of the built-in compilation workflow (see "Makefiles," below).
  • hwname, which lets you provide a name for the job during Makefile flow.
• For SDSoC only:
  • estimate, to use the Xilinx toolchain's resource estimation facility. The job will skip synthesis and execution on the FPGA.
  • directives, which lets you provide the name of a TCL file with a set of HLS directives (pragmas) to use during compilation.
  • platform, the name of the FPGA target to use.
• For SDAccel (F1) only:
  • mode, which lets you choose between software emulation (sw_emu), hardware emulation (hw_emu) and full hardware synthesis (hw).

Use -F <option>=<value> to specify these options with curl.

Viewing Jobs

To see a list of the current jobs, get /jobs.csv:

$ curl $POLYPHEMUS/jobs.csv

To get details about a specific job, request /jobs/<name>:

$ curl $POLYPHEMUS/jobs/d988ruiuAk4

You can also download output files from a job:

$ curl -O $POLYPHEMUS/jobs/d988ruiuAk4/files/code/compiled.o

There is also a JSON list of all the files at /jobs/$ID/files.

Makefiles

Larger projects that use multiple sources and need them to linked in a particular fashion should use the make configuration option. With this option, Polyphemus will run the provided Makefile instead of running its own commands and assume that the artifact is built when the command terminates successfully.

For the job configuration options listed above, Polyphemus provides the additional variables ESTIMATE and DIRECTIVES. Make sure that your Makefile uses these variables to do the appropriate thing during compilation.