Using Sharrow#

This page will walk through an exercise of running a model with sharrow.

How it Works#

Sharrow accelerates ActivitySim in part by using numba to create optimized and pre-compiled versions of utility specification files, and caching those bits of code to disk.

Important

Running the compiler needs to be done in single-process mode, otherwise the various process all do the compiling and compete to write to the same cache location on disk, which is likely to fail. You can safely run in multiprocessing mode after all the compilation for all model components is complete.

Top-Level Activation Options#

Activating sharrow is done at the top level of the model settings file, typically settings.yaml, by setting the sharrow configuration setting to True:

sharrow: True

The default operation for sharrow is to attempt to use the sharrow compiler for all model specifications, and to revert to the legacy pandas-based evaluation if the sharrow compiler encounters a problem. Alternatively, the sharrow setting can also be set to require or test. The require setting will cause the model simply fail if sharrow encounters a problem, which is useful if the user is interested in ensuring maximum performance. The test setting will run the model in a mode where both sharrow and the legacy pandas-based evaluation are run on each model specification, and the results are compared to ensure they are substantially identical. This is useful for debugging and testing, but is not recommended for production runs as it is much slower than running only one evaluation path or the other.

Testing is strongly recommended during model development, as it is possible to write expressions that are valid in one evaluation mode but not the other. This can happen if model data includes NaN values (see Performance Considerations), or when using arithmatic on logical values (see Arithmetic on Logical Values).

Caching of Precompiled Functions#

The first time you run a model with sharrow enabled, the compiler will run and create a cache of compiled functions. This can take a long time, especially for models with many components or complex utility specifications. However, once the cache is created, subsequent runs of the model will be much faster. By default, the cached functions are stored in a subdirectory of the platformdirs.user_cache_dir directory, which is located in a platform-specific location:

Windows: %USERPROFILE%\AppData\Local\ActivitySim\ActivitySim\Cache\...
MacOS: ~/Library/Caches/ActivitySim/...
Linux: ~/.cache/ActivitySim/... or ~/$XDG_CACHE_HOME/ActivitySim/...

The cache directory can be changed from this default location by setting the sharrow_cache_dir setting in the settings.yaml file. Note if you change this setting and provide a relative path, it will be interpreted as relative to the model working directory, and cached functions may not carry over to other model runs unless copied there by the user.

Model Design Requirements#

Activating the sharrow optimizations also requires using the new SkimDataset interface for skims instead of the legacy SkimDict, and internally recoding zones into a zero-based contiguous indexing scheme.

Zero-based Recoding of Zones#

Using sharrow requires recoding zone id’s to be zero-based contiguous index values, at least for internal usage. This recoding needs to be written into the input table list explicitly. For example, the following snippet of a settings.yaml settings file shows the process of recoding zone ids in the input files.

input_table_list:
  - tablename: land_use
    filename: land_use.csv
    index_col: zone_id
    recode_columns:
      zone_id: zero-based
  - tablename: households
    filename: households.csv
    index_col: household_id
    recode_columns:
      home_zone_id: land_use.zone_id

For the land_use table, the zone_id field is marked for recoding explicitly as zero-based, which will turn whatever nominal id’s appear in that column into zero-based index values, as well as store a mapping of the original values that is used to recode and decode zone id’s when used elsewhere.

The first “elsewhere” recoding is in the households input table, where we will map the home_zone_id to the new zone id’s by pointing the recoding instruction at the land_use.zone_id field. If zone id’s appear in other input files, they need to be recoded in those fields as well, using the same process.

The other places where we need to handle zone id’s is in output files. The following snippet of a settings.yaml settings file shows how those id’s are decoded in various output files. Generally, for columns that are fully populated with zone id’s (e.g. tour and trip ends) we can apply the decode_columns settings to reverse the mapping and restore the nominal zone id’s globally for the entire column of data. For columns where there is some missing data flagged by negative values, the “nonnegative” filter is prepended to the instruction.

output_tables:
  action: include
  tables:
    - tablename: land_use
      decode_columns:
        zone_id: land_use.zone_id
    - tablename: households
      decode_columns:
        home_zone_id: land_use.zone_id
    - tablename: persons
      decode_columns:
        home_zone_id: land_use.zone_id
        school_zone_id: nonnegative | land_use.zone_id
        workplace_zone_id: nonnegative | land_use.zone_id
    - tablename: tours
      decode_columns:
        origin: land_use.zone_id
        destination: land_use.zone_id
    - tablename: trips
      decode_columns:
        origin: land_use.zone_id
        destination: land_use.zone_id

Measuring Performance#

Testing with sharrow requires two steps: test mode and production mode.

In test mode, the code is run to compile all the spec files and ascertain whether the functions are working correctly. Test mode is expected to be slow, potentially much slower than older versions of ActivitySim, especially for models with small populations and zone systems, as the compile time is a function of the complexity of the utility functions and not a function of the number of households or zones. Once the compile and test is complete, production mode can then just run the pre-compiled functions with sharrow, which is much faster.

It is possible to run test mode and production mode independently using the existing activitysim run command line tool, pointing that tool to the test and production configurations directories as appropriate.

To generate a meaningful measure of performance enhancement, it is necessary to compare the runtimes in production mode against equivalent runtimes with sharrow disabled. This is facilitated by the activitysim workflow command line tool, which permits the use of pre-made batches of activitysim runs, as well as automatic report generation from the results. For more details on the use of this tool, see workflows.

Sharrow Compatability and Limitations#

In general, utility specification files contain a sequence of :ref:expressions that are evaluated for each row in a main DataFrame. In legacy ActivitySim, there are two fundamental evaluation modes:

pandas.DataFrame.eval, which is the default, and
plain Python eval, which is used when the expression is prefixed with an @ symbol.

Under the pandas.DataFrame.eval mode, expressions are evaluated within the context of the current main dataframe only. References can be made to other columns in that dataframe directly, i.e. if the dataframe contains a column named “income”, the expression can reference that value as “income”. However, the expression can only reference other values that are stored in the current row of the dataframe. The available syntax for these expressions is a subset of regular python, see the supported syntax section of the pandas documentation for details.

Under the plain eval mode, expressions are evaluated within a broader context that potentially includes other variables (which other variables is component dependent) and constants defined in the model settings file. References to columns in the main dataframe must be made indirectly via attribution, i.e. if the dataframe contains a column named “income”, the expression needs to reference that value as “df.income”. Typically, references to skims, time tables, and anything else that isn’t the simple original dataframe use this mode. While it is sometimes not as fast as pandas.DataFrame.eval, this mode is far more flexible, as you can write basically any valid Python expression, including calling other functions, accessing or manipulating table metadata, indexes, or adjacent rows, etc.

Within sharrow, the distinction between these two modes is ignored, as sharrow uses a completely different evaluation system routed through numba. The @ prefix does not need to be stripped or added anywhere, it is simply ignored. The expression can reference other columns of the main dataframe directly or indirectly, so that either “income” or “df.income” is a valid reference to that column in the main dataframe. You can also reference other variables, including skims as usual for each component and constants defined in in the model settings file. Within scheduling models, references can also be made to the timetable as “tt”, accessing that interfaces methods to compute presence and length of adjacent and available time windows. However, you cannot write arbitrary Python code. External functions are not allowed unless they have already been compiled with numba’s njit wrapper. Typically, unless specifically constructed to be allowed for a model component, cross-referencing, merging or reindexing against tables other than the main dataframe is not allowed. Such customizations will generally require a custom extension.

Sharrow only runs for utility specification files. ActivitySim also features the ability to apply pre-processors and post-processors to annotate tables, which use specification files that look very similar to utility specifications. These functions do not run through sharrow, and are a great place to relocate expressions that need to run arbitrary code or join data from other tables.

Temporary Variables#

Temporary variables can be created from @ mode expressions by adding a variable name beginning with an underscore before the @, e.g. _stops_on_leg@df.trip_count-1.

In legacy mode, temporary variables are useful but they can consume substantial memory, as the variable is computed and stored for every row in the entire dataframe before it can be used in other expressions. In sharrow, temporary variables are allocated, used, and freed for each row separately, so no extra memory is required.

Pandas-only Expressions#

In legacy mode, expressions can be evaluated using expressions that tap into the full pandas library, including the ability to call pandas functions and methods directly. This is not possible in sharrow, as the expressions are compiled into numba code, which does not have access to the pandas library. If a pandas function is needed, it must be called in a pre-processor. However, many pandas functions can be replaced with numpy functions, which are available in numba. For example, df.income.fillna(0) can be replaced with np.nan_to_num(df.income).

Switchable Expressions#

As a general rule, it is best to write each utility expression in a manner that is cross-compatible between sharrow and legacy evaluation modes, even if that means transitioning a few expressions or parts of expressions into preprocessors.

However, sometimes this is not possible or makes writing the expressions excessively complex. In this case, it is possible to write a toggling expression, where the individual expression evaluated is different for sharrow and legacy modes. The special comment string # sharrow: splits the expression, with everything before this comment evaluated under the legacy process, and everything after evaluated only when sharrow is enabled.

An example of a switching expression is found in the trip destination utilities found in several examples:

`@np.log1p(size_terms.get(df.alt_dest, df.purpose)) # sharrow: np.log1p(size_terms['sizearray'])`

Here, size_terms is a DataFrameMatrix class instance, a special class written into ActivitySim to facilitate reading from a DataFrame as if it were a 2-d array. As it is a special purpose class written in Python, the numba compiler cannot handle it directly. Fortunately, sharrow provides an alternative: passing the size terms as a xarray DataArray. This has a slightly different interface, so the sharrow and legacy evaluation modes require different expressions. The switching expression is used to handle the DataFrameMatrix on the left (before “# sharrow:”) and the DataArray on the right.

Optional Variables#

In some cases, a variable may be used where it is available, but is not strictly necessary for the model to run. For example, a model may have a reference to mode choice logsums, but the model can still run without them, if it is being used prior to when logsums are calculated. In this case, the variable can be accessed using the get method, which allows for a default value if the variable is not found.

`@df.get('mode_choice_logsum', 0)`

Performance Considerations#

Sharrow is usually expected to bring significant performance gains to ActivitySim. However, the numba engine beneath sharrow has a few known performance limitations, most notably when using strings. To enhance performance, it is best to limit the number of string operations, including simple equality checks (e.g. df.tour_purpose == 'work'). Ideally, such string operations won’t appear in utility specifications at all, or if they do appear, they are executed only once and stored in a temporary value for re-use as needed.

A good approach to reduce string operations in model spec files is to convert string columns to integer or categorical columns in preprocessors. This can be done using the map method, which can be used to convert strings to integers, for example:

`df['fuel_type'].map({'Gas': 1, 'Diesel': 2, 'Hybrid': 3}).fillna(-1).astype(int)`

Alternatively, data columns can be converted to categorical columns with well-defined structures. Recent versions of sharrow have made significant improvements in handling of unordered categorical values, allowing for the use of possibly more intuitive categorical columns. For example, the fuel type column above could instead be redefined as a categorical column with the following code:

`df['fuel_type'].astype(pd.CategoricalDtype(categories=['Gas', 'Diesel', 'Hybrid'], ordered=False))`

It is important that the categories are defined with the same set of values in the same order, as any deviation will from this will void the compiler cache and cause the model specification to be recompiled. This means that using x.astype('category') is not recommended, as the categories will be inferred from the data and may not be consistent across multiple calls to the model specification evaluator.

Note

Beginning with ActivitySim version 1.3, string-valued columns created in preprocessors are converted to categorical columns automatically, which means that ignoring encoding for string-valued outputs is equivalent to using the astype('category') method, and is not recommended.

For models with utility expressions that include a lot of string comparisons, (e.g. because they are built for the legacy pandas.eval interpreter and have not been updated) sharrow can be disabled by setting

compute_settings:
  sharrow_skip: true

in the component’s configuration yaml file.

In addition, by default sharrow also tries to optimize performance by setting the fastmath flag to True in the numba compiler. This makes the compiler generate faster code, by assuming that all variables have finite values (not NaN or Inf). If the model has expressions that use variables that can contains NaN or Inf values, the fastmath flag can be disabled by setting

compute_settings:
  fastmath: false

Multiprocessing Performance#

Sharrow leverages a number of performance enhancing techniques, including parallelization of various computations. This multi-threading can provide significant benefits within a single-process, but if enabled alongside ActivitySim’s multiprocessing paradigm, the multi-threading does more harm than good, as too many threads will compete for limited computational resources. To avoid this, the user should completely disable multi-threading and rely exclusively on multiprocessing to generate parallelism. This can be done by setting a number of thread-limiting environment variables before running Python, or immediately at the start of a Python script before ActivitySim is loaded:

import os
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["OPENBLAS_NUM_THREADS"] = "1"
os.environ["NUMBA_NUM_THREADS"] = "1"
os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"

Limited Tracing and Estimation Mode Capabilities#

When running sharrow-optimized code, large parts of certain calculations are routed through numba to enhance speed and reduce memory footprint. Many intermediate arrays are not created, or are allocated and released by numba in a manner not visible to the general Python interpreter. As a result, in many places the “trace” capabilities of ActivitySim are limited. These capabilities output intermediate calculations in extreme detail and are typically only used in debugging and model development, not in production model runs that need to be optimized for speed. Tracing features can be re-enabled by switching sharrow off or by using “test” mode, which runs both sharrow and legacy evaluations together to confirm they are substantially identical. In this fashion, detailed tracing (albeit slow) remains available for users on all models when needed.

Similar constraints apply to estimation mode, as complete estimation mode output capabilities are not yet integrated with the sharrow engine. Estimation mode remains fully available when running with sharrow disabled.

Arithmetic on Logical Values#

In expressions written in specification files, boolean values must be treated with care. When an expression is evaluated in the legacy implementation, the addition of two boolean values will be processed according to numpy logic, such that:

np.array([True]) + np.array([True]) == np.array([True])
np.array([True]) + np.array([False]) == np.array([True])
np.array([False]) + np.array([True]) == np.array([True])
np.array([False]) + np.array([False]) == np.array([False])

When the same expression is evaluated using sharrow, the expression is evaluated using Pythonesque rules, such that logical values are implicitly upcast to integers, giving:

True + True == 2
True + False == 1
False + True == 1
False + False == 0

If this value is later upcast to a number and used in a mathematical calculation (e.g. multiplied by a float-valued coefficient), obviously the results will vary, as in the first case the result is never other than 1 or 0, but in the latter case the result can also be 2. This mismatch can be readily avoided by wrapping the term in an extra logic gate, which will evaluate the same in both environments:

(True + True)>0 == True
(True + False)>0 == True
(False + True)>0 == True
(False + False)>0 == False

Digital Encoding#

Sharrow is compatible with and able to efficiently use digital encoding. These encodings are applied to data either prospectively (i.e. before ActivitySim ever sees the skim data), or dynamically within a run using the taz_skims.digital-encoding or taz_skims.zarr-digital-encoding settings in the network_los.yaml settings file. The only difference between these two settings is that the former applies this digital encoding internally every time you run the model, while the latter applies it prior to caching encoded skims to disk in Zarr format (and then reuses those values without re-encoding on subsequent runs with the same data). Dictionary encoding (especially joint dictionary encodings) can take a long time to prepare, so caching the values can be useful. As read/write speeds for zarr files are fast, you usually won’t notice a meaningful performance degradation when caching, so the default is generally to use zarr-digital-encoding.

Very often, data can be expressed adequately with far less memory than is needed to store a standard 32-bit floating point representation. There are two simple ways to reduce the memory footprint for data: fixed point encoding, or dictionary encoding.

Fixed Point Encoding#

In fixed point encoding, which is also sometimes called scaled integers, data is multiplied by some factor, then rounded to the nearest integer. The integer is stored in memory instead of a floating point value, but the original value can be (approximately) restored by reversing the process. An offset factor can also be applied, so that the domain of values does not need to start at zero.

For example, instead of storing matrix table values as 32-bit floating point values, they could be multiplied by a scale factor (e.g., 100) and then converted to 16-bit integers. This uses half the RAM and can still express any value (to two decimal point precision) up to positive or negative 327.68. If the lowest values in that range are never needed, it can also be shifted, moving both the bottom and top limits by a fixed amount. Then, for a particular scale $\mu$ and shift $\xi$ (stored in metadata), from any array element $i$ the implied (original) value $x$ can quickly be recovered by evaluating $(i / \mu) - \xi$.

Fixed point digital encoding can be applied to matrix tables in the skims using options in the network_los.yaml settings file. Making transformations currently also requires shifting the data from OMX to ZARR file formats; future versions of ActivitySim may accept digitally encoded data directly from external sources.

To apply the default 16-bit encoding to individual named skim variables in the TAZ skims, just give their names under the zarr-digital-encoding setting like this:

taz_skims:
    omx: skims.omx
    zarr: skims.zarr
    zarr-digital-encoding:
        - name: SOV_TIME
        - name: HOV2_TIME

If some variables can use less RAM and still be represented adequately with only 8-bit integers, you can specify the bitwidth as well:

taz_skims:
    omx: skims.omx
    zarr: skims.zarr
    zarr-digital-encoding:
        - name: SOV_TIME
        - name: HOV2_TIME
        - name: SOV_TOLL
          bitwidth: 8
        - name: HOV2_TOLL
          bitwidth: 8

If groups of similarly named variables should have the same encoding applied, they can be identified by regular expressions (“regex”) instead of explicitly giving each name. For example:

taz_skims:
    omx: skims.omx
    zarr: skims.zarr
    zarr-digital-encoding:
        - regex: .*_TIME
        - regex: .*_TOLL
          bitwidth: 8

Dictionary Encoding#

For dictionary encoding, a limited number of unique values are stored in a lookup array, and then each encoded value is stored as the position of the value (or its closest approximation) in the lookup array. If there are fewer than 256 unique values, this can allow the storage of those values to any level of precision (even float64 if needed) while using only a single byte per array element, plus a small fixed amount of overhead for the dictionary itself. The overhead memory doesn’t scale with the dimensions of the array, so this works particularly well for models with thousands of zones.

Dictionary encoding can be applied to a single variable in a similar fashion as fixed point encoding, giving the dictionary bit width in the by_dict setting, or as an additional setting value.

taz_skims:
    omx: skims.omx
    zarr: skims.zarr
    zarr-digital-encoding:
        - name: TRANSIT_FARE
          by_dict: 8
        - name: TRANSIT_XFERS
          by_dict: true
          bitwidth: 8

The most dramatic memory savings can be found when the categorical correlation (also known as Cramér’s V) between multiple variables is high. In this case, we can encode more than one matrix table using the same dictionary lookup indexes. There may be some duplication in the lookup table, (e.g. if FARE and XFER are joint encoded, and if a FARE of 2.25 can be matched with either 0 or 1 XFER, the 2.25 would appear twice in the lookup array for FARE, once for each value of XFER.)

Since it is the lookup indexes that scale with the number of zones and consume most of the memory for large zone systems, putting multiple variables together into one set of indexes can save a ton of memory, so long as the overhead of the lookup array does not combinatorially explode (hence the need for categorical correlation).

Practical testing for large zone systems suggest this method of encoding can reduce the footprint of some low variance data tables (especially transit data) by 95% or more.

Applying joint dictionary encoding requires more than one variable name, so only the regex form works here. Use wildcards to match on name patterns, or select a few specific names by joining them with the pipe operator (|).

taz_skims:
    omx: skims.omx
    zarr: skims.zarr
    zarr-digital-encoding:
        - regex: .*_FARE|.*_WAIT|.*_XFERS
          joint_dict: true
        - regex: FERRYTIME|FERRYFARE|FERRYWAIT
          joint_dict: true

For more details on all the settings available for digital encoding, see DigitalEncoding.

Troubleshooting#

If you encounter errors when running the model with sharrow enabled, it is important to address them before using the model for analysis. This is especially important when errors are found running in “test” mode (activated by sharrow: test in the top level settings.yaml). Errors may indicate that either sharrow or the legacy evaluator is not correctly processing the mathematical expressions in the utility specifications.

“utility not aligned” Error#

One common error that can occur when running the model with sharrow in “test” mode is the “utility not aligned” error. This error occurs when a sharrow compiled utility calculation does not sufficiently match the legacy utility calculation. We say “sufficiently” here because the two calculations may have slight differences due to numerical precision optimizations applied by sharrow. These optimizations can result in minor differences in the final utility values, which are typically inconsequential for model results. However, if the differences are too large, the “utility not aligned” error will be raised. This error does not indicate whether the incorrect result is from the sharrow or legacy calculation (or both), and it is up to the user to determine how to align the calculations so they are reflective of the model developer’s intent.

To troubleshoot the “utility not aligned” error, the user can use a Python debugger to compare the utility values calculated by sharrow and the legacy evaluator. ActivitySim also includes error handler code that will attempt to find the problematic utility expression and print it to the console or log file, under the heading “possible problematic expressions”. This can be helpful in quickly narrowing down which lines of a specification file are causing the error.

Common causes of the “utility not aligned” error include:

model data includes NaN values but the component settings do not disable fastmath (see Performance Considerations)
incorrect use of arithmatic on logical values (see Arithmetic on Logical Values)

Insufficient system resources#

For large models run on large servers, it is possible to overwhelm the system with too many processes and threads, which can result in the following error:

OSError: Insufficient system resources exist to complete the requested service

This error can be resolved by reducing the number of processes and/or threads per process. See Multiprocessing and Multithreading in the User’s Guide for more information on how to adjust these settings.

Permission Error#

If running a model using multiprocessing with sharrow enabled, it is necessary to have pre-compiled all the utility specifications to prevent the multiple processes from competing to write to the same cache location on disk. Failure to do this can result in a permission error, as some processes may be unable to write to the cache location.

PermissionError: The process cannot access the file because it is being used by another process

To resolve this error, run the model with sharrow enabled in single-process mode to pre-compile all the utility specifications. If that does not resolve the error, it is possible that some compiling is being triggered in multiprocess steps that is not being handled in the single process mode. This is likely due to the presence of string or categorical columns created in a preprocessor that are not being stored in a stable data format. To resolve this error, ensure that all expressions in pre-processors are written in a manner that results in stable data types (e.g. integers, floats, or categorical columns with a fixed set of categories). See see Performance Considerations) for examples.