The Legate project endeavors to democratize computing by making it possible for all programmers to leverage the power of large clusters of CPUs and GPUs by running the same code that runs on a desktop or a laptop at scale. Using this technology, computational and data scientists can develop and test programs on moderately sized data sets on local machines and then immediately scale up to larger data sets deployed on many nodes in the cloud or on a supercomputer without any code modifications.
The Legate project is built upon two foundational principles:
- For end users, such as computational and data scientists, the programming model must be identical to programming a single sequential CPU on their laptop or desktop. All concerns relating to parallelism, data distribution, and synchronization must be implicit. The cloud or a supercomputer should appear as nothing more than a super-powerful CPU core.
- Software must be compositional and not just interoperable (i.e. functionally correct). Libraries developed in the Legate ecosystem must be able to exchange partitioned and distributed data without requiring "shuffles" or unnecessary blocking synchronization. Computations from different libraries should be able to use arbitrary data and still be reordered across abstraction boundaries to hide communication and synchronization latencies where the original sequential semantics of the program allow. This is essential for achieving speed-of-light performance on large scale machines.
The Legate project is still in its nascent stages of development, but much of the fundamental architecture is in place. We encourage development and contributions to existing Legate libraries, as well as the development of new Legate libraries. Pull requests are welcomed.
If you have questions, please contact us at legate(at)nvidia.com.
- Legate
Computational problems today continue to grow both in their complexity as well as the scale of the data that they consume and generate. This is true both in traditional HPC domains as well as enterprise data analytics cases. Consequently, more and more users truly need the power of large clusters of both CPUs and GPUs to address their computational problems. Not everyone has the time or resources required to learn and deploy the advanced programming models and tools needed to target this class of hardware today. Legate aims to bridge this gap so that any programmer can run code on any scale machine without needing to be an expert in parallel programming and distributed systems, thereby allowing developers to bring the problem-solving power of large machines to bear on more kinds of challenging problems than ever before.
The Legate Core is our version of Apache Arrow. Apache Arrow has significantly improved composability of software libraries by making it possible for different libraries to share in-memory buffers of data without unnecessary copying. However, it falls short when it comes to meeting two of our primary requirements for Legate:
- Arrow only provides an API for describing a physical representation of data as a single memory allocation. There is no interface for describing cases where data has been partitioned and then capturing the logical relationships of those partitioned subsets of data.
- Arrow is mute on the subject of synchronization. Accelerators such as GPUs achieve significantly higher performance when computations are performed asynchronously with respect to other components of the system. When data is passed between libraries today, accelerators must be pessimistically synchronized to ensure that data dependences are satisfied across abstraction boundaries. This might result in tolerable overheads for single GPU systems, but can result in catastrophically poor performance when hundreds of GPUs are involved.
The Legate Core provides an API very similar to Arrow's interface with several important distinctions that provide stronger guarantees about data coherence and synchronization to aid library developers when building Legate libraries. These guarantees are the crux of how libraries in the Legate ecosystem are able to provide excellent composability.
The Legate Core API imports several important concepts from Arrow such that
users that are familiar with Arrow already will find it unsurprising. We use
the same type system representation as Arrow so libraries that have already
adopted it do not need to learn or adapt to a new type system. We also reuse
the concept of an Array
from Arrow. The LegateArray class supports many of the same methods as
the Arrow Array interface (we'll continue to add methods to improve
compatibility). The main difference is that instead of obtaining
Buffer
objects from arrays to describe allocations of data that back the array, the
Legate Core API introduces a new primitive called a LegateStore which
provides a new interface for reasoning about partitioned and distributed
data in asynchronous execution environments.
Any implementation of a LegateStore must maintain the following guarantees
to clients of the Legate Core API (i.e. Legate library developers):
- The coherence of data contained in a
LegateStoremust be implicitly managed by the implementation of the Legate Core API. This means that no matter where data is requested to perform a computation in a machine, the most recent modifications to that data in program order must be reflected. It should never be clients responsibility to maintain this coherence. - It should be possible to create arbitrary views onto
LegateStoreobjects such that library developers can precisely describe the working sets of their computations. Modifications to views must be reflected onto all aliasing views data. This property must be maintained by the Legate Core API implementation such that it is never the concern of clients. - Dependence management between uses of the
LegateStoreobjects and their views is the responsibility of Legate Core API regardless of what (asynchronous) computations are performed onLegateStoreobjects or their views. This dependence analysis must be both sound and precise. It is illegal to over-approximate dependences. This dependence analysis must also be performed globally in scope. Any use of theLegateStoreon any processor/node in the system must abide by the original sequential semantics of the program
Note that we do not specify exactly what the abstractions are that are needed
for implementing LegateStore objects. Our goal is not prescribe what these
abstractions are as they may be implementation dependent. Our only requirements
are that they have these properties to ensure that incentives are aligned in
such a way for Legate libraries to achieve a high degree of composability
at any scale of machine. Indeed, these requirements shift many of the burdens
that make implementing distributed and accelerated libraries hard off of the
library developers and onto the implementation of the Legate Core API. This
is by design as it allows the costs to be amortized across all libraries in
the ecosystem and ensures that Legate library developers are more productive.
Our implementation of the Legate Core API is built on top of the Legion programming model and runtime system. Legion was originally designed for large HPC applications that target supercomputers and consequently applications written in the Legion programming model tend to both perform and scale well on large clusters of both CPUs and GPUs. Legion programs are also easy to port to new machines as they inherently decouple the machine-independent specification of computations from decisions about how that application is mapped to the target machine. Due to this abstract nature, many programmers find writing Legion programs challenging. By implementing the Legate Core API on top of Legion, we've made it easier to use Legion such that developers can still get access to the benefits of Legion without needing to learn all of the lowest-level interfaces.
The Legion programming model
greatly aids in implementing the Legate Core API. Data types from libraries,
such as arrays in cuNumeric are mapped down onto LegateStore objects
that wrap Legion data types such as logical regions or futures.
In the case of regions, Legate application libraries rely heavily on
Legion's support for partitioning of logical regions into arbitrary
subregion views.
Each library has its own heuristics for computing such partitions that
take into consideration the computations that will access the data, the
ideal sizes of data to be consumed by different processor kinds, and
the available number of processors. Legion automatically manages the coherence
of subregion views regardless of the scale of the machine.
Computations in Legate application libraries are described by Legion tasks.
Tasks describe their data usage in terms of LegateStore objects, thereby
allowing Legion to infer where dependences exist. Legion uses distributed
bounding volume hierarchies, similar to a high performance ray-tracer,
to soundly and precisely perform dependence analysis on logical regions
and insert the necessary synchronization between tasks to maintain the
original sequential semantics of a Legate program.
Each Legate application library also comes with its own custom Legion mapper that uses heuristics to determine the best choice of mapping for tasks (e.g. are they best run on a CPU or a GPU). All Legate tasks are currently implemented in native C or CUDA in order to achieve excellent performance on the target processor kind, but Legion has bindings in other languages such as Python, Fortran, and Lua for users that would prefer to use them. Importantly, by using Legion, Legate is able to control the placement of data in order to leave it in-place in fast memories like GPU framebuffers across tasks.
When running on large clusters, Legate leverages a novel technology provided by Legion called "control replication" to avoid the sequential bottleneck of having one node farm out work to all the nodes in the cluster. With control replication, Legate will actually replicate the Legate program and run it across all the nodes of the machine at the same time. These copies of the program all cooperate logically to appear to execute as one program. When communication is necessary between different computations, the Legion runtime's program analysis will automatically detect it and insert the necessary data movement and synchronization across nodes (or GPU framebuffers). This is the transformation that allows sequential programs to run efficiently at scale across large clusters as though they are running on a single processor.
Legate Core is available on conda:
conda install -c nvidia -c conda-forge -c legate legate-core
Only linux-64 packages are available at the moment.
The default package contains GPU support, and is compatible with CUDA >= 11.8
(CUDA driver version >= r520), and Volta or later GPU architectures. There are
also CPU-only packages available, and will be automatically selected by conda
when installing on a machine without GPUs.
See BUILD.md for instructions on building Legate Core from source.
After installing the Legate Core library, the next step is to install a Legate
application library such as cuNumeric. The installation process for a
Legate application library will require you to provide a pointer to the location
of your Legate Core library installation as this will be used to configure the
installation of the Legate application library. After you finish installing any
Legate application libraries, you can then simply replace their import statements
with the equivalent ones from any Legate application libraries you have installed.
For example, you can change this:
import numpy as npto this:
import cunumeric as npAfter this, you can use the legate driver script in the bin directory of
your installation to run any Python program.
You can also use the standard Python interpreter, but in that case configuration options can only be passed through the environment (see below), and some options are not available (check the output of legate --help for more details).
For example, to run your script in the default configuration (4 CPUs cores and 4 GB of memory) just run:
$ legate my_python_program.py [other args]
The legate script also allows you to control the amount of resources that
Legate consumes when running on the machine. The --cpus and --gpus flags
are used to specify how many CPU and GPU processors should be used on a node.
The --sysmem flag can be used to specify how many MBs of DRAM Legate is allowed
to use per node, while the --fbmem flag controls how many MBs of framebuffer
memory Legate is allowed to use per GPU. For example, when running on a DGX
station, you might run your application as follows:
$ legate --cpus 16 --gpus 4 --sysmem 100000 --fbmem 15000 my_python_program.py
This will make 16 CPU processors and all 4 GPUs available for use by Legate.
It will also allow Legate to consume up to 100 GB of DRAM memory and 15 GB of
framebuffer memory per GPU for a total of 60 GB of GPU framebuffer memory. Note
that you probably will not be able to make all the resources of the machine
available for Legate as some will be used by the system or Legate itself for
meta-work. Currently if you try to exceed these resources during execution then
Legate will inform you that it had insufficient resources to complete the job
given its current mapping heuristics. If you believe the job should fit within
the assigned resources please let us know so we can improve our mapping heuristics.
There are many other flags available for use in the legate driver script that you
can use to communicate how Legate should view the available machine resources.
You can see a list of them by running:
$ legate --help
In addition to running Legate programs, you can also use Legate in an interactive
mode by simply not passing any *py files on the command line. You can still
request resources just as you would though with a normal file. Legate will
still use all the resources available to it, including doing multi-node execution.
$ legate --cpus 16 --gpus 4 --sysmem 100000 --fbmem 15000
Welcome to Legion Python interactive console
>>>
Note that Legate does not currently support multi-tenancy cases where different users are attempting to use the same hardware concurrently.
As a convenience, several command-line options can have their default values set via environment variables. These environment variables, their corresponding command- line options, and their default values are as follows.
| CLI Option | Env. Variable | Default Value |
|---|---|---|
| --omps | LEGATE_OMP_PROCS | 0 |
| --ompthreads | LEGATE_OMP_THREADS | 4 |
| --utility | LEGATE_UTILITY_CORES | 2 |
| --sysmem | LEGATE_SYSMEM | 4000 |
| --numamem | LEGATE_NUMAMEM | 0 |
| --fbmem | LEGATE_FBMEM | 4000 |
| --zcmem | LEGATE_ZCMEM | 32 |
| --regmem | LEGATE_REGMEM | 0 |
| --eager-alloc-percentage | LEGATE_EAGER_ALLOC_PERCENTAGE | 50 |
If Legate is compiled with networking support (see the
installation section),
it can be run in parallel by using the --nodes option followed by the number of nodes
to be used. Whenever the --nodes option is used, Legate will be launched
using mpirun, even with --nodes 1. Without the --nodes option, no launcher will
be used. Legate currently supports mpirun, srun, and jsrun as launchers and we
are open to adding additional launcher kinds. You can select the
target kind of launcher with --launcher.
Legate also comes with several tools that you can use to better understand
your program both from a correctness and a performance standpoint. For
correctness, Legate has facilities for constructing both dataflow
and event graphs for the actual run of an application. These graphs require
that you have an installation of GraphViz
available on your machine. To generate a dataflow graph for your Legate
program simply pass the --dataflow flag to the legate.py script and after
your run is complete we will generate a dataflow_legate.pdf file containing
the dataflow graph of your program. To generate a corresponding event graph
you simply need to pass the --event flag to the legate.py script to generate
a event_graph_legate.pdf file. These files can grow to be fairly large for non-trivial
programs so we encourage you to keep your programs small when using these
visualizations or invest in a robust PDF viewer.
For profiling all you need to do is pass the --profile flag to Legate and
afterwards you will have a legate_prof directory containing a web page that
can be viewed in any web browser that displays a timeline of your program's
execution. You simply need to load the index.html page from a browser. You
may have to enable local JavaScript execution if you are viewing the page from
your local machine (depending on your browser).
We recommend that you do not mix debugging and profiling in the same run as some of the logging for the debugging features requires significant file I/O that can adversely effect the performance of the application.
Same as normal Python programs, Legate programs can be run using Jupyter Notebook. Currently we support single node execution with multiple CPUs and GPUs, and plan to support multi-node execution in the future. We leverage Legion's Jupyter support, so you may want to refer to the relevant section in Legion's README. To simplify the installation, we provide a script specifically for Legate libraries.
Please install Legate, then run the following command to install a default Jupyter kernel:
legate-jupyter
If installation is successful, you will see some output like the following:
Jupyter kernel spec Legate_SM_GPU (Legate_SM_GPU) has been installed
Legate_SM_GPU is the default kernel name.
You will need to start a Jupyter server, then you can use a Jupyter notebook from any browser. Please refer to the following two sections from the README of the Legion Jupyter Notebook extension
- Start the Jupyter Notebook server
- Use the Jupyter Notebook in the browser
The Legate Jupyter kernel is configured according to the command line arguments
provided at install time. Standard legate options for Core, Memory, and
Mult-node configuration may be provided, as well as a name for the kernel:
legate-jupyter --name legate_cpus_2 --cpus 2
Other configuration options can be seen by using the --help command line option.
We provide a Jupyter magic command to display the IPython kernel configuration.
%load_ext legate.jupyter
%legate_info
results in output:
Kernel 'Legate_SM_GPU' configured for 1 node(s)
Cores:
CPUs to use per rank : 4
GPUs to use per rank : 0
OpenMP groups to use per rank : 0
Threads per OpenMP group : 4
Utility processors per rank : 2
Memory:
DRAM memory per rank (in MBs) : 4000
DRAM memory per NUMA domain per rank (in MBs) : 0
Framebuffer memory per GPU (in MBs) : 4000
Zero-copy memory per rank (in MBs) : 32
Registered CPU-side pinned memory per rank (in MBs) : 0
-
Does Legate only work on NVIDIA hardware? No, Legate will run on any processor supported by Legion (e.g. x86, ARM, and PowerPC CPUs), and any network supported by GASNet or UCX (e.g. Infiniband, Cray, Omnipath, and (ROC-)Ethernet based interconnects).
-
What languages does the Legate Core API have bindings for? Currently the Legate Core bindings are only available in Python. Watch this space for new language bindings soon or make a pull request to contribute your own. Legion has a C API which should make it easy to develop bindings in any language with a foreign function interface.
-
Do I have to build drop-in replacement libraries? No! While we've chosen to provide drop-in replacement libraries for popular Python libraries to illustrate the benefits of Legate, you are both welcomed and encouraged to develop your own libraries on top of Legate. We promise that they will compose well with other existing Legate libraries.
-
What other libraries are you planning to release for the Legate ecosystem? We're still working on that. If you have thoughts about what is important please let us know so that we can get a feel for where best to put our time.
-
Can I use Legate with other Legion libraries? Yes! If you're willing to extract the Legion primitives from the
LegateStoreobjects you should be able to pass them into other Legion libraries such as FlexFlow. -
Does Legate interoperate with X? Yes, probably, but we don't recommend it. Our motivation for building Legate is to provide the bare minimum subset of functionality that we believe is essential for building truly composable software that can still run at scale. No other systems out there met our needs. Providing interoperability with those other systems will destroy the very essence of what Legate is and significantly dilute its benefits. All that being said, Legion does provide some means of doing stop-the-world exchanges with other runtime system running concurrently in the same processes. If you are interested in pursuing this approach please open an issue on the Legion github issue tracker as it will be almost entirely orthogonal to how you use Legate.
See the discussion of contributing in CONTRIBUTING.md.
A complete list of available features can is found in the Legate Core documentation.
We recommend starting by experimenting with at least one Legate application library to test out performance and see how Legate works. If you are interested in building your own Legate application library, we recommend that you investigate our Legate Hello World application library that provides a small example of how to get started developing your own drop-in replacement library on top of Legion using the Legate Core library.