Getting Performance With Intel® SDC

Compiled code does not necessarily faster than interpreted one. There are several key elements impacting whether compiled code will outperform an interpreted one or not. These are compilation overheads, overheads related to converting Python objects to native structures and back, amount of parallelism in compiled code, to what extent the code is “static”, and many other factors. The following section will discuss each of these aspects in greater detail.

Compilation Overheads

Compilation of the code may lead to significantly better performance, however, it comes with some costs. First, there is a cost associated with the need to compile a code. Depending on the size of compiled function and many other factors the compilation time may be material.

With above in mind there are several general recommendations when using Intel SDC will be beneficial.

1. Compile application hotspots only.

   • Focus on where application spends most time.

2. Compile parts of the code where parallelism resides.

   • Remember that Intel SDC has been created to extract parallelism. If there is no parallelism in the code, then it unlikely to perform visibly better than interpreted code.

3. Compilation overheads quickly become negligible when compiled function is called multiple times.

   • If arguments to compiled function are type stable then Numba* won’t recompile the function on each call.

4. Use Numba* caching for ahead-of-time compilation.

5. Minimize the number of columns in dataframes in a region being compiled.

   • Intel SDC performs dataframe analysis at compilation time, as a result the compilation may take significantly longer if the number of columns is large. Among key features to reduce the number of columns in compile time analysis is to read in memory only columns actually used in computations.

Boxing And Unboxing Overheads

All Python* objects are accessed through GIL, the Global Interpreter Lock, a mutex that protects Python objects preventing multiple threads from executing bytecodes concurrently. While it is okay technology for interpreted code it kills any sort of parallelism. For Intel SDC and Numba* to be able to extract parallelism, the GIL must be released. To release GIL Python objects must be converted into native structures that do not require GIL. The implication is additional overheads to convert input arguments into native objects (unboxing) and to convert results back to Python objects (boxing).

In most cases these overheads are reasonably small as Intel SDC and Numba try to avoid extensive memory copy during boxing and unboxing. However, sometimes memory copy is unavoidable, and passing large dataframe as input or as a returning value may be expensive.

It is almost always a good idea to originate large amounts of data from within compiled region. For example, reading large CSV file from outside compiled region and then passing that large dataframe as input may result in visible unboxing costs.

The rule of thumb is “Generate and process massive data from within the compiled region. Pass through JIT region borders small data only”.

Automatic Parallelization

Intel SDC parallelizes most of Pandas* operations so that users do not typically need to take extra steps besides using @njit decorator. However, sometimes you might want to extract additional parallelism available in a JIT-region. To do that you need to add parallel=True option to @njit decorator:

Example 2: Parallelizing Pandas* Workflow

import pandas as pd
from numba import njit, prange

# Dataset for analysis
FNAME = "employees.csv"


# This function gets compiled by Numba* and multi-threaded
@njit(parallel=True)
def get_analyzed_data():
    df = pd.read_csv(FNAME)
    s_bonus = pd.Series(df['Bonus %'])
    s_first_name = pd.Series(df['First Name'])

    # Use explicit loop to compute the mean. It will be compiled as parallel loop
    m = 0.0
    for i in prange(s_bonus.size):
        m += s_bonus.values[i]
    m /= s_bonus.size

    names = s_first_name.sort_values()
    return m, names


# Printing names and their average bonus percent
mean_bonus, sorted_first_names = get_analyzed_data()
print(sorted_first_names)
print('Average Bonus %:', mean_bonus)

Other Performance Tips

Please refer to Numba Performance Tips for additional hints on obtaining good performance with Intel SDC.