Llama is a tool for running UNIX commands inside of AWS Lambda. Its goal is to make it easy to outsource compute-heavy tasks to Lambda, with its enormous available parallelism, from your shell.

Most notably, llama includes llamacc, a drop-in replacement for gcc or clang which executes the compilation in the cloud, allowing for considerable speedups building large C or C++ software projects.

Lambda offers nearly-arbitrary parallelism and burst capacity for compute, making it, in principle, well-suited as a backend for interactive tasks that briefly require large amounts of compute. This idea has been explored in the ExCamera and gg papers, but is not widely accessible at present.

Here are a few performance results from my testing demonstrating the current speedups achievable from llamacc:

As you can see, Llama is capable of speedups for large builds even on my large, powerful desktop system, and the advantage is more pronounced on smaller workstations.

• A Linux x86_64 machine. Llama only supports that platform for now. Cross-compilation should in theory be possible but is not implemented.
• The Go compiler. Llama is tested on v1.16 but older versions may work.
• An AWS account

You'll need to install Llama from source. You can run

go install github.com/nelhage/llama/cmd/...@latest

or clone this repository and run

go install ./...

If you want to build C++, you'll want to symlink llamac++ to point at llamacc:

ln -nsf llamacc "$(dirname $(which llamacc))/llamac++"

Llama needs access to your AWS credentials. You can provide them in the environment via AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY, but the recommended approach is to use ~/.aws/credentials, as used by. Llama will read keys out of either.

The account whose credentials you use must have sufficient permissions. The following should suffice:

• AmazonEC2ContainerRegistryFullAccess
• AmazonS3FullAccess
• AWSCloudFormationFullAccess
• AWSLambda_FullAccess
• IAMFullAccess

Llama includes a CloudFormation template and a command which uses it to bootstrap all required resources. You can read the template to see what it's going to do.

Once your AWS credentials are ready, run

$ llama bootstrap

to create the required AWS resources. By default, it will prompt you for an AWS region to use; you can avoid the prompt using (e.g.) llama -region us-west-2 bootstrap.

If you get an error like

Creating cloudformation stack...
Stack created. Polling until completion...
Stack is in rollback: ROLLBACK_IN_PROGRESS. Something went wrong.
Stack status reason: The following resource(s) failed to create: [Repository, Bucket]. Rollback requested by user.

then you can go to the AWS web console, and find the relevant CloudFormation stack. The event log should have more useful errors explaining what went wrong. You will then need to delete the stack before retrying the bootstrap.

You'll need to build a container with an appropriate version of GCC for llamacc to use.

If you are running Debian or Ubuntu, you can use scripts/build-gcc-image to automatically build a Debian image and Lambda function matching your local system:

$ scripts/build-gcc-image

If you want more control or are running another distribution, you can look at images/gcc-focal for an example Dockerfile to build a compiler package. You can build that or a similar image into a Lambda function using llama update-function like so:

$ llama update-function --create --build=images/gcc-focal gcc

To use llamacc, run a build using make or a similar build system with a much higher -j concurrency than you normally would -- try 5-10x the number of local cores,, and using llamacc or llamac++ as your compiler. For example, you might invoke

$ make -j100 CC=llamacc CXX=llamac++

llamacc takes a number of configuration options from the environment, so that they're easy to pass through your build system. The currently supported options include.

It is strongly recommended that you use absolute paths if you set LLAMACC_LOCAL_CC and LLAMACC_LOCAL_CXX. Not all build systems will preserve $PATH all the way down to llamacc, so if you don't use absolute paths, you can get build failures that are difficult to diagnose.

You can use llama invoke to execute individual commands inside of Lambda. The syntax is llama invoke <function> <command> args.... <function> must be the name of a Lambda function using the Llama runtime. So, for instance, we can inspect the OS running inside our Lambda image:

$ llama invoke gcc uname -a
Linux 169.254.248.253 4.14.225-175.364.amzn2.x86_64 #1 SMP Mon Mar 22 22:06:01 UTC 2021 x86_64 x86_64 x86_64 GNU/Linux

If your function consumes files as input or output, you can use the -f and -o options to specify that files should be passed between the local and remote nodes. For instance:

$ llama invoke -f README.md:INPUT -o OUTPUT gcc sh -c 'sha256sum INPUT > OUTPUT'; cat OUTPUT
16c399c108bb783fc5c4529df4fecd0decb81bc0707096ebd981ab2b669fae20  INPUT

Note the use of LOCAL:REMOTE syntax to optionally specify different paths between the local and remote ends.

llama xargs provides an xargs-like interface for running commands in parallel in Lambda. Here's an example:

The optipng command compresses PNG files and otherwise optimizes them to be as small as possible, typically used in order to save bandwidth and speed load times on image assets. optipng is somewhat computationally expensive and compressing a large number of PNG files can be a slow operation. With llama, we can optimize a large of images by outsourcing the computation to lambda.

I prepared a directory full of 151 PNG images of the original Pokémon, and benchmarked how long it took to optimize them using 8 concurrent processes on my desktop:

$ time ls -1 *.png | parallel -j 8 optipng {} -out optimized/{/}
[...]
real    0m45.090s
user    5m33.745s
sys     0m0.924s

Once we've prepared and optipng lambda function (we'll talk about setup in a later section), we can use llama to run the same computation in AWS Lambda:

$ time ls -1 *.png | llama xargs -logs -j 151 optipng optipng '{{.I .Line}}' -out '{{.O (printf "optimized/%s" .Line)}}'
real    0m16.024s
user    0m2.013s
sys     0m0.569s

We use llama xargs, which works a bit like xargs(1), but runs each input line as a separate command in Lambda. It also uses the Go template language to provide flexibility in substitutions, and offers the special .Input and .Output methods (.I and .O for short) to mark files to be passed back and forth between the local environment and Lambda.

Lambda's CPUs are slower than my desktop and the network operations have overhead, and so we don't see anywhere near a full 151/8 speedup. However, the additional parallelism still nets us a 3x improvement in real-time latency. Note also the vastly decreased user time, demonstrating that the CPU-intensive work has been offloaded, freeing up local compute resources for interactive applications or other use cases.

This operation consumed about 700 CPU-seconds in Lambda. I configured optipng to have 1792MB of memory, which is the point at which lambda allocates a full vCPU to the process. That comes out to about 1254400 MB-seconds of usage, or about $0.017 assuming I'm already out of the Lambda free tier.

The llama runtime is designed to make it easy to bridge arbitrary images into Lambda. You can look at images/optipng/Dockerfile in this repository for a well-commented example explaining how you can wrap an arbitrary image inside of Lambda for use by Llama.

Once you have a Dockerfile or a Docker image, you can use llama update-function to upload it to ECR and manage the associated Lambda function. For instance, we could build optipng for the above example like so:

$ llama update-function --create --build=images/optipng optipng

When specifying the memory size for your functions, note that Lambda assigns CPU resources to functions based on their memory allocation. At 1,769 MB, your function will have the equivalent of one full core.

Llama is in large part inspired by gg, a tool for outsourcing builds to Lambda. Llama is a much simpler tool but shares some of the same ideas and is inspired by a very similar vision of using Lambda as high-concurrency burst computation for interactive uses.