Distributed Nested Performance Contexts in SQL
A toolkit for analysing performance data from various components when running large workflows with Parsl
pip install .
# Remove existing DB
$ rm dnpc.sqlite3
# Import a runinfo directory
$ python3 -m dnpcsql.import_parsl_runinfo ~/parsl/src/parsl/runinfo
# See which span types have been imported
$ python3 -m dnpcsql.list_span_types
parsl.monitoring.workflow
parsl.monitoring.task
parsl.monitoring.try
workqueue.task
workqueue.worker
# List most common sequence of events
$ python3 -m dnpcsql.list_event_sequences
... verbose output ...
-
An SQLite database with a basic schema
-
Importers which take data from various components and populate the database.
Examples of importers:
- Parsl monitoring database
- Work Queue transaction log files
- High Throughput Executor worker log files
- application-specific data
Hopefully it is straightforward to add new importers for new components, sometimes for permanent uses, sometimes for ad-hoc measurements of some particular feature.
-
Analysis code
Python and SQL code which makes interesting reports from the data in the database. Some of this code will be for specific components, but some reports can be generic across all components.
Some of this code is structured as command line tools.
Some is in Python libraries that is intended to be used in user's scripts or notebooks.
The basic object that is being analysed is a "span". This term comes from the distributed tracing/observability community. For example, LightStep defines a span as:
"A span is a named, timed operation that represents a piece of the workflow."
In the case of parsl, spans include things like: parsl level tasks, parsl level tries, entire parsl workflows, executor level execution attempts, batch provider blocks, executor level task executors, task executions on a particular worker.
A span will have multiple events associated with it. For the purposes of dnpcsql, an event has a timestamp and a type. Usually there is at least a start and end event. Often the end event will describe how the span ended - for example, a failure event or successful completion event.
At the application level above parsl, there might be spans that represent the entire execution of an application, and each individual component of an application workflow; and inside the application code that runs on a worker, that execution might be divided into spans.
Spans can be nested inside other spans (the N in dnpcsql). Often spans will be nested in a way that is "almost" a 1-1 relationship: for example, very often an application level task will correspond with exactly one parsl task, which will correspond with exactly one parsl try, which will correspond with exactly one executor level task execution. But that correspondence is not exact: what happens if a task is never executed? or if an application level task fails before it is even submitted to parsl?
Spans can also be associated with other spans as "facets" of each other: for example, the parsl monitoring database and an in-development parsl performance event system independently present views of the "same" things, such as parsl tasks. Neither is a parent to each other.
An flow for performance data might look like this:
-
stuff happens in a workflow (that we will want to analyse the performance of) and events are logged in ad-hoc formats
-
dnpcsql users import revelant information about spans, from various sources, into the database
-
users get specific data from the database using SQL queries
-
users use matplotlib to plot that data
-
users examine plots
The query/plot/examine stages are intended to be quite dynamic, with users writing and modifying their own code to explore performance data, rather than a small number of pre-defined plots being made by developers and released to dnpcsql users. Code in the dnpcrepo around analysis should look more like helper code, rather than pre-packaged analyses.
Users are free to structure their database schema as they please. There are three different kinds of structure that have been used with dnpcsql so far:
-
span / event / subspan / facet - one table represents all spans, regardless of type; another represents events in those spans; and two more represent subspan/superspan and facet relationships. This is span/event type unaware.
-
Type-aware tables - eg. parsl monitoring.db, with a table for tasks, another table for tries, etc. Different columns in the database represent timestamps for different events (and other data about the types of span)
-
raw json log records - for example, from cloudwatch. In this model, SQL JSON functions are used in queries to access relevant fields at index and query time, rather than at import time. This leads to more complex queries, but also allows access of the entirety of a log record.
Data in all of these forms (and other forms) can exist in the sqlite3 database at once, accessed with SQL queries that understand all of the forms that they are querying.
Different components identify their spans differently, and with different scopes.
For example, the parsl monitoring database identifies a try span, globally, with a three part composite key: run ID, task ID and try ID. So when joining tries to other tables, a three part join needs to happen.
Work queue log files, however, for a particular run (the lifetime of a manager?) use a single integer key to identify tasks, scoped only within the life of that manager. There is no further attempt to give the manager a global identity in the same way that parsl gives runs a global run ID.
This can be awkward when trying to import two data sources independently with the intention of joining them later: something involving surrogate keys probably needs to happen. For example, in the parsl case, a hierarchical import using some kind of surrogate key that models that certain work queue log files live inside the rundir of a particular identified parsl run.
There are lots of common queries such as "tell me information about the distribution of durations between a start and end event" which can be plotted in (for example) matplotlib.
This project should provide some helpers to make those plots, which take output from an arbitrary SQL query that outputs in a known format (such as one row for each span, with two columns: the first event and the second event).
Arranging the data into these structures happens in SQL, which for many queries is the optimal place to do it; then plotting happens in Python around matplotlib.
See dnpcsql/twoevents.py for an example of this.
This project originates from work with the realtime monitoring component of Parsl and with trying to understand performance of larger systems where parsl is not the only major component.
As a complement to Parsl's realtime, dashboard oriented monitoring system, this project is aimed at people who are comfortable writing their own analyses of performance data (which is often the case as Parsl users tend to be scientific programmers); it is intended to support after-the-fact, non-realtime performance analyses, from multiple data sources (primarily log files of different components, the sqlite3 database of parsl monitoring information, but also for example, data coming from AWS cloudwatch as part of a funcX side-project)
... or areas that other people might be interested in poking at.
-
Non-SQL databases and query languages - for example, graph databases and graph query languages. There's nothing here that is intimately tied to SQL; for example, the first iteration of this project was an entirely in-memory Python object collection.
-
realtime database update as a workflow progresses - experience with Parsl has shown this to be expensive, and in the case of post-hoc performance, an expense that doesn't need to be paid.
-
enforcing a common reporting format from components - this project has to deal with data as it comes from the components (for example, a big log file per run, a parsl monitoring database containing many workflows data, many separate log files one per core), and it is the job of component specific importers to understand these formats.
-
using a richer standardised data model (standard span or event types, for example) for all components
-
figuring out that your clocks are not synchronised across machines
https://docs.google.com/presentation/d/1wW2FLYMTeko1_Qxhx8IB3T3DtXedTb-rT38mrB2qkiw/edit?usp=sharing
- Presentation to #parsl-hackers slack channel about this project, April 2023
http://netlogger.lbl.gov/ - NetLogger, a very similar project. Especially see NetLogger's logging best practices document: https://docs.google.com/document/d/1oeW_l_YgQbR-C_7R2cKl6eYBT5N4WSMbvz0AT6hYDvA/edit
https://github.com/dwreeves/dbt_linreg - linear regression in SQL - this is an example of using templating/libraries to make an SQL/hybrid DSL, which might be an interesting path forwards for query writing.