Arkouda nightly performance charts
Mike Merrill's SIAM PP-22 Talk
Bill Reus' March 2021 talk at the NJIT Data Science Seminar
Bill Reus' CHIUW 2020 Keynote video and slides
Mike Merrill's CHIUW 2019 talk
Exploratory data analysis (EDA) is a prerequisite for all data science, as illustrated by the ubiquity of Jupyter notebooks, the preferred interface for EDA among data scientists. The operations involved in exploring and transforming the data are often at least as computationally intensive as downstream applications (e.g. machine learning algorithms), and as datasets grow, so does the need for HPC-enabled EDA. However, the inherently interactive and open-ended nature of EDA does not mesh well with current HPC usage models. Meanwhile, several existing projects from outside the traditional HPC space attempt to combine interactivity and distributed computation using programming paradigms and tools from cloud computing, but none of these projects have come close to meeting our needs for high-performance EDA.
To fill this gap, we have developed a software package, called Arkouda, which allows a user to interactively issue massively parallel computations on distributed data using functions and syntax that mimic NumPy, the underlying computational library used in the vast majority of Python data science workflows. The computational heart of Arkouda is a Chapel interpreter that accepts a pre-defined set of commands from a client (currently implemented in Python) and uses Chapel's built-in machinery for multi-locale and multithreaded execution. Arkouda has benefited greatly from Chapel's distinctive features and has also helped guide the development of the language.
In early applications, users of Arkouda have tended to iterate rapidly between multi-node execution with Arkouda and single-node analysis in Python, relying on Arkouda to filter a large dataset down to a smaller collection suitable for analysis in Python, and then feeding the results back into Arkouda computations on the full dataset. This paradigm has already proved very fruitful for EDA. Our goal is to enable users to progress seamlessly from EDA to specialized algorithms by making Arkouda an integration point for HPC implementations of expensive kernels like FFTs, sparse linear algebra, and graph traversal. With Arkouda serving the role of a shell, a data scientist could explore, prepare, and call optimized HPC libraries on massive datasets, all within the same interactive session.
Arkouda is not trying to replace Pandas but to allow for some Pandas-style operation at a much larger scale. In our experience Pandas can handle dataframes up to about 500 million rows before performance becomes a real issue, this is provided that you run on a sufficiently capable compute server. Arkouda breaks the shared memory paradigm and scales its operations to dataframes with over 200 billion rows, maybe even a trillion. In practice we have run Arkouda server operations on columns of one trillion elements running on 512 compute nodes. This yielded a >20TB dataframe in Arkouda.
- Prerequisites
- Building Arkouda
- Testing Arkouda
- Running arkouda_server
- Logging
- Type Checking in Arkouda
- Versioning
- External Systems Integration
- Metrics
- Contributing
Prerequisites toc
For a complete list of requirements for Arkouda, please review REQUIREMENTS.md.
For detailed prerequisite information and installation guides, please review the install guide for your operating system.
Building Arkouda toc
In order to run the Arkouda server, it must first be compiled. Detailed instructions on the build process can be found at BUILD.md.
Testing Arkouda toc
(click to see more)
There are two unit test suites for Arkouda, one for Python and one for Chapel. As mentioned above, the Arkouda
Python test harness leverages multiple libraries such as pytest and
pytest-env that must be installed via pip3 install -e .[dev],
whereas the Chapel test harness does not require any external librares.
The default Arkouda test executes the Python test harness and is invoked as follows:
make testThe Chapel unit tests can be executed as follows:
make test-chapelBoth the Python and Chapel unit tests are executed as follows:
make test-allFor more details regarding Arkouda testing, please consult the Python test README and Chapel test README, respectively.
Running arkouda_server toc
The command-line invocation depends on whether you built a single-locale version (with CHPL_COMM=none) or
multi-locale version (with CHPL_COMM set to the desired number of locales).
Single-locale startup:
./arkouda_serverMulti-locale startup (user selects the number of locales):
./arkouda_server -nl 2Memory tracking is turned on by default now, you can run server with memory tracking turned off by
./arkouda_server --memTrack=falseBy default, the server listens on port 5555. This value can be overridden with the command-line flag
--ServerPort=1234
Trace logging messages are turned on by default and turned off by using the --trace=false flag
Other command line options are available and can be viewed by using the --help flag
./arkouda-server --helpWith the addition of two server startup flags, --autoShutdown and --serverInfoNoSplash, running the arkouda_server from a script is easier than ever.
To connect to the server via a script, you'll first have to issue a subprocess command to start the arkouda_server with the optional configuration flags.
import subprocess
# Update the below path to point to your arkouda_server
cmd = "/Users/<username>/Documents/git/arkouda/arkouda_server -nl 1 --serverInfoNoSplash=true --autoShutdown=true"
p = subprocess.Popen(cmd, stdout=subprocess.PIPE, shell=True)This will allow you to access the server output, which on launch using --serverInfoNoSplash=true will be a JSON string with the server configuration which can be parsed for the server host, port, and other potentially useful information.
For a full example and explanation, view the Running From Script document.
Sanity check arkouda_server toc
To sanity check the arkouda server, you can run
make checkThis will start the server, run a few computations, and shut the server down. In addition, the check script can be executed against a running server by running the following Python command:
python3 tests/check.py localhost 5555Token-Based Authentication in Arkouda toc
Arkouda features a token-based authentication mechanism analogous to Jupyter, where a randomized alphanumeric string is generated or loaded at arkouda_server startup. The command to start arkouda_server with token authentication is as follows:
./arkouda_server --authenticateThe generated token is saved to the tokens.txt file which is contained in the .arkouda directory located in the same working directory the arkouda_server is launched from. The arkouda_server will re-use the same token until the .arkouda/tokens.txt file is removed, which forces arkouda_server to generate a new token and corresponding tokens.txt file.
In situations where a user-specified token string is preferred, this can be specified in the ARKOUDA_SERVER_TOKEN environment variable. As is the case with an Arkouda-generated token, the user-supplied token is saved to the .arkouda/tokens.txt file for re-use.
Setting Per-Locale Memory and CPU Core Limits toc
By default, each Arkouda locale utilizes all available memory and CPU cores on the host machine. However, it is possible to set per-locale limits for both memory as well as CPU cores.
There are three approaches to setting the max memory used by each Arkouda locale. Firstly, the built-in Chapel approach sets the max per-locale memory to an explicit number of bytes via the --memMax startup parameter. For example, to set the max memory utilized by each locale to 100 GB, the Arkouda startup command would include the following:
./arkouda_server --memMax=100000000000
The Arkouda dynamic memory limit approach sets the per-locale memory limit based upon a configurable percentage of available memory on each locale host. Prior to the execution of each command, the MemoryMgmt localeMemAvailable function does the following on each locale:
- Verifies the projected, additional per-locale memory required by the incoming command does not exceed the memory currently allocated to Arkouda. If the projected, additional memory is within the memory currently allocated to Arkouda on each locale, the command is allowed to proceed.
- If the projected, additional per-locale memory exceeds the memory currently allocated to Arkouda on any locale, localeMemAvailable checks if the configurable percentage of available memory on each node will accommodate the projected, additional memory of the incoming command. If so, the command is allowed to proceed.
- If the projected, additional per-locale memory required by the incoming command exceeds the configured percentage of available memory on any locale, localeMemAvailable returns false and a corresponding error is thrown in the ServerConfig overMemLimit function.
In the example below, dynamic memory checking is enabled with the default availableMemoryPct of 90, configuring Arkouda to throw an error if (1) the projected, additional memory required for a command exceeds memory currently allocated to Arkouda on 1..n locales and (2) the projected, additional memory will exceed 90 percent of available memory on 1..n locales.
./arkouda_server --MemoryMgmt.memMgmtType=MemMgmtType.DYNAMIC
Setting additionalMemoryPct to 70 would result in the following startup command:
./arkouda_server --MemoryMgmt.memMgmtType=MemMgmtType.DYNAMIC ----MemoryMgmt.additionalMemoryPct=70
Important note: dynamic memory checking works on Linux and Unix systems only.
In the final, default approach, the max memory utilized by each locale is set as percentage of physical memory on the locale0 host, defaulting to 90 percent. If another percentage is desired, this is set via the --perLocaleMemLimit startup parameter. For example, to set max memory utilized by each locale to seventy (70) percent of physical memory on locale0, the Arkouda startup command would include the following:
./arkouda_server --perLocaleMemLimit=70
The max number of CPU cores utilized by each locale is set via the CHPL_RT_NUM_THREADS_PER_LOCALE environment variable. An example below sets the maximum number of cores for each locale to 16:
export CHPL_RT_NUM_THREADS_PER_LOCALE=16
Connecting to Arkouda toc
The client connects to the arkouda_server either by supplying a host and port or by providing a connect_url connect string:
arkouda.connect(server='localhost', port=5555)
arkouda.connect(connect_url='tcp://localhost:5555')When arkouda_server is launched in authentication-enabled mode, clients connect by either specifying the access_token parameter or by adding the token to the end of the connect_url connect string:
arkouda.connect(server='localhost', port=5555, access_token='dcxCQntDQllquOsBNjBp99Pu7r3wDJn')
arkouda.connect(connect_url='tcp://localhost:5555?token=dcxCQntDQllquOsBNjBp99Pu7r3wDJn')Note: once a client has successfully connected to an authentication-enabled arkouda_server, the token is cached in the user's $ARKOUDA_HOME .arkouda/tokens.txt file. As long as the arkouda_server token remains the same, the user can connect without specifying the token via the access_token parameter or token url argument.
Logging toc
The Arkouda server features a Chapel logging framework that prints out the module name, function name and line number for all logged messages. An example is shown below:
2021-04-15:06:22:59 [ConcatenateMsg] concatenateMsg Line 193 DEBUG [Chapel] creating pdarray id_4 of type Int64
2021-04-15:06:22:59 [ServerConfig] overMemLimit Line 175 INFO [Chapel] memory high watermark = 44720 memory limit = 30923764531
2021-04-15:06:22:59 [MultiTypeSymbolTable] addEntry Line 127 DEBUG [Chapel] adding symbol: id_4
Available logging levels are ERROR, CRITICAL, WARN, INFO, and DEBUG. The default logging level is INFO where all messages at the ERROR, CRITICAL, WARN, and INFO levels are printed. The log level can be set globally by passing in the --logLevel parameter upon arkouda_server startup. For example, passing the --logLevel=LogLevel.DEBUG parameter as shown below sets the global log level to DEBUG:
./arkouda_server --logLevel=LogLevel.DEBUG
In addition to setting the global logging level, the logging level for individual Arkouda modules can also be configured. For example, to set MsgProcessing to DEBUG for the purposes of debugging Arkouda array creation, pass the MsgProcessing.logLevel=LogLevel.DEBUG parameter upon arkouda_server startup as shown below:
./arkouda_server --MsgProcessing.logLevel=LogLevel.DEBUG --logLevel=LogLevel.WARN
In this example, the logging level for all other Arkouda modules will be set to the global value WARN.
Arkouda logs can be written either to the console (default) or to the arkouda.log file located in the .arkouda directory. To enable log output to the arkouda.log file, start Arkouda as follows with the --logChannel flag:
./arkouda_server --logChannel=LogChannel.FILE
All incoming Arkouda server commands submitted by the Arkouda client can be logged to the commands.log file located in the .arkouda directory. Arkouda command logging is enabled as follows:
./arkouda_server --logCommands=true
The Arkouda command logging capability has a variety of uses, one of which is replaying analytic or data processing scenarios in either interactive or batch mode. Moreover, a sequence of Arkouda server commands provides the possibility of utilizing Arkouda clients developed in other languages such as Rust or Go. In still another use case, command logging in Arkouda provides a command sequence for starting Arkouda via cron job and processing large amounts of data into Arkouda arrays or dataframes, thereby obviating the need for a user to wait for well-known data processing/analysis steps to complete; this use case is of particular value in situations where the data loading process is particularly time-intensive. Finally, command logging provides a means of integrating a non-interactive Arkouda data processing/analysis sequence into a data science workflow implemented in a framework such as Argo Workflows or Kubeflow.
Type Checking in Arkouda toc
Both static and runtime type checking are becoming increasingly popular in Python, especially for large Python code bases such as those found at dropbox. Arkouda uses mypy for static type checking and typeguard for runtime type checking.
(click to see more)
Enabling runtime as well as static type checking in Python starts with adding type hints, as shown below to a method signature:
def connect(server : str="localhost", port : int=5555, timeout : int=0,
access_token : str=None, connect_url=None) -> None:
mypy static type checking can be invoked either directly via the mypy command or via make:
$ mypy arkouda
Success: no issues found in 16 source files
$ make mypy
python3 -m mypy arkouda
Success: no issues found in 16 source files
Runtime type checking is enabled at the Python method level by annotating the method if interest with the @typechecked decorator, an example of which is shown below:
@typechecked
def save(self, prefix_path : str, dataset : str='array', mode : str='truncate') -> str:
Type checking in Arkouda is implemented on an "opt-in" basis. Accordingly, Arkouda continues to support duck typing for parts of the Arkouda API where type checking is too confining to be useful. As detailed above, both runtime and static type checking require type hints. Consequently, to opt-out of type checking, simply leave type hints out of any method declarations where duck typing is desired.
Versioning toc
Beginning after tag v2019.12.10 versioning is now performed using Versioneer
which determines the version based on the location in git.
An example using a hypothetical tag 1.2.3.4
git checkout 1.2.3.4
python -m arkouda |tail -n 2
>> Client Version: 1.2.3.4
>> 1.2.3.4
# If you were to make uncommitted changes and repeat the command you might see something like:
python -m arkouda|tail -n 2
>> Client Version: 1.2.3.4+0.g9dca4c8.dirty
>> 1.2.3.4+0.g9dca4c8.dirty
# If you commit those changes you would see something like
python -m arkouda|tail -n 2
>> Client Version: 1.2.3.4+1.g9dca4c8
>> 1.2.3.4+1.g9dca4c8In the hypothetical cases above Versioneer tells you the version and how far / how many commits beyond the tag your repo is.
When building the server-side code the same versioning information is included in the build. If the server and client do not match you will receive a warning. For developers this is a useful reminder when you switch branches and forget to rebuild.
# Starting the arkouda when built from tag 1.2.3.4 shows the following in the startup banner
arkouda server version = 1.2.3.4
# If you built from an arbitrary branch the version string is based on the derived coordinates from the "closest" tag
arkouda server version = v2019.12.10+1679.abc2f48a
# The .dirty extension denotes a build from uncommitted changes, or a "dirty branch" in git vernacular
arkouda server version = v2019.12.10+1679.abc2f48a.dirtyFor maintainers, creating a new version is as simple as creating a tag in the repository; i.e.
git checkout master
git tag 1.2.3.4
python -m arkouda |tail -n 2
>> Client Version: 1.2.3.4
>> 1.2.3.4
git push --tagsExternal Systems Integration toc
Integrating Arkouda with cloud environments enables users to access Arkouda from machine learning (ML) and deep learning (DL) workflows deployed to Kubernetes as an example. Detailed discussions regarding Arkouda systems integration and specific instructions for registering/deregistering Arkouda with Kubernetes are located in EXTERNAL INTEGRATION.md
Metrics toc
Arkouda provides a separate, dedicated zmq socket to enable generation and export of a variety of system, locale, user, and request metrics. Arkouda generated metrics in a format compatible with Prometheus, Grafana, and TimescaleDB. An Arkouda Prometheus exporter that serves as a Prometheus scrape target will be made available soon in the arkouda-contrib repository. A detailed discussion of Arkouda metrics is located in METRICS.md
Contributing to Arkouda toc
If you'd like to contribute, we'd love to have you! Before jumping in and adding issues or writing code, please see CONTRIBUTING.md.