Craig Weber

Benchmarking Go and Python

Sometimes I’m curious about the performance of different languages. At work, I usually write Python, but I often find tasks that are inherently parallelizable and could thus benefit from parallel execution. Unfortunately, Python is notoriously difficult to parallelize1. In one case, we needed to validate that a table of values of a particular type could be convertible into a values of a different type based on some known set of conversion rules. Since Go is a great language for writing concurrent programs (and executing them in parallel), I decided to compare a sequential Python implementation to sequential and parallel Go implementations.

To simplify the problem for benchmarking, I decided to constrain the input type to strings and the output type to integers, and the “conversion rule” is strconv.Atoi() in Go and int(string) in Python. In all examples, I’m reading comma-separated string values from stdin, building a 2D string array in memory, and then timing the conversion of strings to ints. I didn’t want to include reading from stdin in my timing, because that would make my benchmarks dependent upon the performance of the data source (a network, the file system, etc). Reading from memory is much less variable.

In Python, the code looks like this:

def validate_rows(rows, col_size):
    for row_id, row in enumerate(rows):
        if len(row) != col_size:
            msg = "Row {} has {} cells, but expected {}\n"
            print(msg.format(row_id, len(row), col_size))
            continue
        for col_id, cell in enumerate(row):
            try:
                int(cell)
            except ValueError as e:
                print("Err at ({}, {}): {}".format(col_id, row_id, e))

And in Go:

func validateRows(rows [][]string, colSize int) {
    for rowID, row := range rows {
        if len(row) != colSize {
            msg := "Row %d has %d cells, but expected %d\n"
            fmt.Fprintf(os.Stderr, msg, rowID, len(row), colSize)
            continue
        }
        for colID, cell := range row {
            if _, err := strconv.Atoi(cell); err != nil {
                msg := "Err at (%d, %d): %v\n"
                fmt.Fprintf(os.Stderr, msg, colID, rowID, err)
            }
        }
    }
}

I also created a simple program that deterministically generates pseudo-random integer data in 2D CSV format and writes it to stdout:

$ go run csvgen.go
USAGE: csvgen <col-count> <row-count>

This lets me pipe data into the Go and Python versions, like so2:

# Sequential Python; 10,000 rows, 1,000 columns
$ go run csvgen.go 1000 10000 | python3 sequential.py
Beginning validation...
Validated 10000 rows of 1000 cells in 0:00:02.990360

# Sequential Go; 10,000 rows, 1,000 columns
$ go run csvgen.go 1000 10000 | go run sequential.go
Beginning validation...
Validated 10000 rows of 1000 cells in 664.804988ms

The Go program was about 4.5 times faster than the Python program. Running this example over and over again on my MacBook Pro produces consistent results, with the Go program outperforming the Python variant by 4-6 times.

But what about parallelism? I’m not at all familiar with parallelism in Python, so I didn’t attempt it (perhaps a reader could supply an implementation, and contact me on Twitter?), but Go makes parallelism very easy, so I decided to give it a shot:

func validateParallel(rows [][]string, coreCount int) {
    wg := sync.WaitGroup{}
    wg.Add(coreCount) // Add `coreCount` goroutines to the WaitGroup
    
    // divide `rows` into `coreCount` blocks of rows, and then dispatch a
    // goroutine to process each block.
    for _, block := range subslice(rows, coreCount) {
        // Create a new variable exclusively for the goroutine that corresponds
        // to this loop iteration. All goroutines can't share one variable,
        // because the variable will be pointing to the last block returned by
        // subslice() before the first goroutine is kicked off, meaning all
        // goroutines would be operating on the last block and the previous
        // blocks would be ignored.
        block := block
        
        go func() {
            validateRows(block, len(rows[0]))
            wg.Done() // signal that this goroutine has finished execution
        }()
    }
    
    wg.Wait() // block until `wg.Done()` has been called `coreCount` times
}

This version takes a 2D input string array and breaks it into coreCount subarrays, where coreCount is intended to be the number of cores on the machine (although it can be any number between 1 and len(rows)). It then dispatches one goroutine (a lightweight, cooperative thread) per subarray, which invokes the validateRows() function from the sequential Go code snippet above. It also waits for each goroutine to finish before returning. Here’s how the function is called (without timing or print statements):

    // query the machine's CPU core count
    coreCount := runtime.NumCPU()
    
    // allow the Go runtime to spin up (at most) `coreCount` threads
    runtime.GOMAXPROCS(coreCount)
    
    // pass the input rows and coreCount to validateParallel()
    validateParallel(rows, coreCount)

This version does about twice as well as the sequential Go algorithm for all inputs I tested:

$ go run csvgen.go 1000 10000 | go run parallel.go
GOMAXPROCS: 4
Beginning validation...
Validated 10000 rows of 1000 cells in 299.069099ms

The performance ratio holds even if we bump up the input volume by an order of magnitude:

# Sequential Python; 100,000 rows, 1,000 columns
$ csvgen 1000 100000 | python3 sequential.py
Beginning validation...
Validated 100000 rows of 1000 cells in 0:00:30.148827

# Sequential Go; 100,000 rows, 1,000 columns
$ csvgen 1000 100000 | go run sequential.go
Beginning validation...
Validated 100000 rows of 1000 cells in 6.700995661s

# Parallel Go; 100,000 rows, 1,000 columns
$ csvgen 1000 100000 | go run parallel.go
GOMAXPROCS: 4
Beginning validation...
Validated 100000 rows of 1000 cells in 2.985182403s

Besides performance, one of the nice things about the Go implementations is that they’re not much less readable than the Python implementation. In fact, I’ve found that the prevalence of types and the absence of features (no inheritance, decorators, exceptions, etc) make Go very quick to learn and easy to reason about.

Feel free to try it yourself, or perhaps improve on my implementations. The source code for this case study can be found here. Share your results (or any other thoughts you had on this post) with me on Twitter.


  1. I got a lot of flack for this from a lot of Python folks who insisted that parallelism in Python was easy, but the fastest parallel Python implementation provided to me was still twice as slow as the sequential Python implementation. I’ve updated my benchmark repo to include the 3 parallel Python implementations.

  2. Several folks were confused about whether or not the compilation of the Go programs (which take place during go run) were included in the benchmarks. They are not; the Go and Python programs are all responsible for timing themselves. This is because I wanted to be able to start time after each program had loaded its input data into a 2D-array in memory, so the I/O wasn’t included in the benchmarks.