Skip to content

Latest commit

 

History

History
284 lines (191 loc) · 21.4 KB

PERFORMANCE.md

File metadata and controls

284 lines (191 loc) · 21.4 KB
Raw

Performance Tuning

Index! Index! Index!

Indexing your CSV files is key for performance. Here's why:

  1. Faster Slicing: slice uses the index to directly retrieve relevant rows.

  2. Instant Row Counts: count the total number of rows instantly. It also sets up the --progressbar faster, and speeds qsv when it needs the row count (which it does for just about every command).

  3. Parallel Processing: Indexing enables multithreading, dramatically speeding up supported commands like stats, frequency, sample, split and tojsonl.

  4. Random Access: The luau command gains random access capabilities. extsort CSV mode requires an index.

  5. Low Overhead: Creating an index is fast and efficient, even for very large files. The million row, 41-column, 520mb NYC 311 benchmark file for instance, takes all of 466 ms to index.

Even if you're only handling a CSV file once, and its not reference data, indexing still makes sense if you're slicing, counting, sampling, using the --progressbar or compiling summary statistics with the stats and frequency commands.

The only time indexing is not useful is when the CSV file is too small to benefit from indexing or when the file is not seekable (e.g. stdin).

To enable automatic indexing:

  • Set the QSV_AUTOINDEX_SIZE environment variable
  • Specify the minimum file size (in bytes) for auto-indexing
# automatically create an index for files larger than 10MB
export QSV_AUTOINDEX_SIZE=10000000

Stats Cache

stats is the primary reason qsv was created. Several projects we were working on required GUARANTEED data type inferences at speed when we first working on it in 2021. As we iterated and started additional projects, we started needing additional capabilities to enable the "automagical metadata" inferencing workflow we wanted for our data ingestion pipelines.

From the original 11 summary statistics in xsv (type, sum, min/max, min/max length, mean, stddev, median, mode & cardinality ), 22 more were added incrementally over time (is_ascii, range, sort_order, sum_length, avg_length, mean_length, sem, variance, cv, nullcount, max_precision, sparsity, mad, lower outer/inner fence, q1, q2_median, q3, iqr, upper inner/outer fence, skewness, mode_count, mode_occurrences, antimode, antimode_count, antimode_occurrences). Check the Wiki for more info.

And some of these stats were relatively expensive to compute, so qsv started caching statistics so it didn't need to recompute them if a file hasn't changed (as most of the files we were working on were historical data).

Slowly, over time, we realized that the cached stats can be used to make other commands faster and smarter - thus the stats cache was born!

  • frequency uses the stats cache to short-circuit compiling frequency tables for ID columns (all unique values) by looking at the cardinality of a column.
  • schema uses the cache to create a JSON Schema Validation file. It uses the cache to set the data type, enum values, const values, minLength, maxLength, minimum and maximum properties in the JSON Schema file.
  • tojsonl uses the cache to set the JSON data type, and to infer boolean JSON properties.
  • sqlp and joinp uses the cache to create a Polars Schema, short-circuting Polars' schema inferencing - which is not as reliable as it depends on sampling the first N rows of a CSV, which may lead to wrong type inferences if the sample size is not large enough (which if set too large, slows down the Polars engine). As the data type inferences of stats are guaranteed, its not only faster, it works all the time!

For the most part, the default caching behavior works transparently, though you will notice several files with the same file stem will start appearing in the same location as your CSV files. As metadata is tiny by nature and very useful on its own, a conscious decision was made not to hide them.

If you want to fine-tune qsv's caching behavior, use the --cache-threshold option. It's one of the few options that based on its value, can be different units:

  • when greater than 1, the threshold in MILLISECONDS before caching stats results

  • when set to 0 - suppresses caching

  • when set to 1 - forces caching

    As stats is much faster with an index. It also controls auto-indexing:

  • when set to a negative number, automatically creates an index when the input file size is greater than the absolute of the provided values in BYTES. The stats cache remains after stats finishes.

  • when set to a negative number AND the number ends with 5, it will automatically create an index, compile the stats, AND then delete the index as well as the stats cache files afterwards.

CPU Optimization

Modern CPUs have various features that the Rust compiler can take advantage of to increase performance. If you want the compiler to take advantage of these CPU-specific speed-ups, set this environment variable BEFORE installing/compiling qsv:

On Linux and macOS:

export CARGO_BUILD_RUSTFLAGS='-C target-cpu=native'

On Windows Powershell:

$env:CARGO_BUILD_RUSTFLAGS='-C target-cpu=native'

Do note though that the resulting binary will only run on machines with the same architecture as the machine you installed/compiled from. To find out your CPU architecture and other valid values for target-cpu:

rustc --print target-cpus

# to find out what CPU features are used by the Rust compiler WITHOUT specifying target-cpu
rustc --print cfg | grep -i target_feature

# to find out what additional CPU features will be used by the Rust compiler when you specify target-cpu=native
rustc --print cfg -C target-cpu=native | grep -i target_feature

# to get a short explanation of each CPU target-feature
rustc --print target-features

Memory Management

Memory Allocator

qsv supports three memory allocators: mimalloc, jemalloc and the standard allocator.

By default, qsv uses mimalloc, a performance-oriented allocator from Microsoft.

You can also use another alternative - the jemalloc allocator, which is the default Linux allocator used by the Pola.rs project for its python bindings, as benchmarks have shown that it performs better than mimalloc on some platforms.

If you don't want to use mimalloc, use the --no-default-features flag when installing/compiling qsv, e.g.:

To use jemalloc

cargo install qsv --path . --no-default-features --features all_features,jemallocator

or

cargo build --release --no-default-features --features all_features,jemallocator

To use the standard allocator

cargo install qsv --path . --no-default-features --features all_features

or

cargo build --release --no-default-features --features all_features

To find out what memory allocator qsv is using, run qsv --version. After the qsv version number, the allocator used is displayed ("standard", "mimalloc" or "jemalloc"). Note that mimalloc is not supported on the x86_64-pc-windows-gnu and arm targets, and you'll need to use the "standard" allocator on those platforms.

Out-of-Memory (OOM) Prevention

Most qsv commands use a "streaming" approach to processing CSVs - "streaming" in the input record-by-record while processing it. This allows it to process arbitrarily large CSVs with constant memory.

There are a number of commands/modes however (denoted by the "exploding head" emoji - 🤯), that require qsv to load the entire CSV into memory - dedup (when not using the --sorted option), reverse, sort, stats (when calculating the "non-streaming" extended stats), table and transpose (when not running in --multipass mode).

NOTE: Though not as flexible, dedup and sort have corresponding "external" versions - extdedup and extsort respectively, that use external memory (i.e. disk) to process arbitrarily large CSVs.

In addition, frequency, schema and tojsonl - though they do not load the entire file into memory, uses additional memory proportional to the cardinality (number of unique values) of each column compared to other "streaming" commands (denoted by the "persevering face" emoji - 😣).

For very large files, this can be a problem, as qsv will run out of memory and panic. To prevent this, qsv has two memory check heuristics when running "non-streaming" commands:

NORMAL mode

  1. at startup, get the TOTAL memory of the system
  2. if the size of the CSV file is greater than TOTAL memory - HEADROOM (default: 20%), qsv will abort with an error

CONSERVATIVE mode

  1. at startup, compute total available memory by adding the current available memory and free swap space
  2. subtract a percentage headroom from the total available memory (default: 20%)
  3. if this adjusted total available memory is less than the size of the CSV file, qsv will abort with an error

The percentage headroom can be changed by setting the QSV_MEMORY_HEADROOM_PCT environment variable to a value between 10 and 90 (default: 20).

This CONSERVATIVE heuristic can have false positives however, as modern operating systems can do a fair bit of juggling to handle file sizes larger than what this heuristic will allow, as it dynamically swaps apps to the swapfile, expand the swapfile, compress memory, etc.

For example, on a 16gb Mac mini running several common apps, it only allowed ~3gb csv files, but in practice, it was able to handle files up to 8gb before this heuristic was added.

To apply this CONSERVATIVE heuristic, you can use the command's --memcheck option or set the QSV_MEMORY_CHECK environment variable.

Otherwise, the default memory check heuristic (NORMAL mode) will only check if the input file's size is larger than the TOTAL memory of the computer minus QSV_MEMORY_HEADROOM_PCT. We still do this to prevent OOM panics, but it's not as restrictive as the CONSERVATIVE heuristic. (e.g. if you have a 16gb computer, the maximum input file size is 12.8gb file - 16gb minus 20% headroom).

NOTE: These memory checks are not invoked when using stdin as input, as the size of the input file is not known. Though schema and tojsonl will still abort if stdin is too large per this memory check as it creates a temporary file from stdin before inferring the schema.

Buffer size

Depending on your filesystem's configuration (e.g. block size, file system type, writing to remote file systems (e.g. sshfs, efs, nfs), SSD or rotating magnetic disks, etc.), you can also fine-tune qsv's read/write buffers.

By default, the read buffer size is set to 128k, you can change it by setting the environment variable QSV_RDR_BUFFER_CAPACITY in bytes.

The same is true with the write buffer (default: 256k) with the QSV_WTR_BUFFER_CAPACITY environment variable.

Multithreading

Several commands support multithreading - count, stats, frequency, sample, schema, split and tojsonl (when an index is available); apply, applydp, dedup, diff, excel, extsort, joinp, snappy, sort, sqlp, to and validate (no index required).

qsv will automatically spawn parallel jobs equal to the detected number of logical processors. Should you want to manually override this, use the --jobs command-line option or the QSV_MAX_JOBS environment variable.

To find out your jobs setting, call qsv --version.

Version details

The --version option shows a lot of information about qsv. It displays:

  • qsv version
  • the memory allocator (standard, mimalloc or jemalloc)
  • all enabled features (apply, fetch, foreach, luau, polars, python, self_update & to)
  • Python version linked if the python feature was enabled
  • Luau version embedded if the luau feature was enabled
  • the number of processors to use for multi-threading commands
  • the number of logical processors detected
  • memory-related OOM prevention info (max "non-streaming" input file size, free swap memory, available memory & total memory)
  • the target platform
  • the Rust version used to compile qsv
  • QSV_KIND - prebuilt, prebuilt-*, installed & compiled. The prebuilts are the qsv binaries published on Github with every release. prebuilt is built using the current Rust stable at the time of release. prebuilt-nightly is built using Rust nightly that passes all CI tests at the time of release. installed is qsv built using cargo install. compiled is qsv built using cargo build.
$ qsv --version
qsv 0.122.0-mimalloc-apply;fetch;foreach;Luau 0.606;python-3.11.0 (v3.11.0:deaf509e8f, Oct 24 2022, 14:43:23) [Clang 13.0.0 (clang-1300.0.29.30)];to;self_update-8-8;3.66 GiB-913.00 MiB-3.69 GiB-16.00 GiB (aarch64-apple-darwin compiled with Rust 1.75.0) compiled

Shows that I'm running qsv version 0.122.0, with the mimalloc allocator (instead of standard or jemalloc), and I have the apply, fetch, foreach, luau, python, to and self_update features enabled, with the exact version of the embedded Luau interpreter, and the python version qsv is dynamically linked against.

It shows qsv will use 8 logical processors out of 8 detected when running multithreaded commands.

It also shows that I can have a maximum input file size of 3.66 GiB for "non-streaming" commands (see Memory Management for more info), 913.00 MiB of free swap memory, 3.69 GiB of available memory and 16.00 GiB of total memory.

The qsv binary was built to target the aarch64-apple-darwin platform (Apple Silicon), compiled using Rust 1.75.0. The binary was compiled using cargo build.

Caching

qsv employs several caching strategies to improve performance:

  • qsv has large read and write buffers to minimize disk I/O. The default read buffer size is 128k and the default write buffer size is 512k. These can be fine-tuned with the QSV_RDR_BUFFER_CAPACITY and QSV_WTR_BUFFER_CAPACITY environment variables.
  • The stats command caches its results in both CSV and JSONL formats. It does this to avoid re-computing the same statistics when the same input file/parameters are used, but also, as statistics are used in several other commands (currently - frequency, schema and tojsonl, with more commands using cached statistics in the future).
    The stats cache are automatically refreshed when the input file is modified the next time the stats command is run or when cache-aware commands attempt to use them. The stats cache is stored in the same directory as the input file. The stats cache files are named with the same file stem as the input file with the stats.csv, stats.csv.json and stats.csv.data.jsonl extensions. The CSV contains the cached stats, the JSON file contains metadata about how the stats were compiled, and the JSONL file is the JSONL version of the stats that can be directly loaded into memory by other commands. The JSONL is used by the frequency, schema and tojsonl commands and will only be generated when the --stats-jsonl option is set.
  • The geocode command memoizes otherwise expensive geocoding operations and will report its cache hit rate. geocode memoization, however, is not persistent across sessions.
  • The fetch and fetchpost commands also memoizes expensive REST API calls. When the --redis option is enabled, it effectively has a persistent cache as the default time-to-live (TTL) before a Redis cache entry is expired is 28 days and Redis entries are persisted across restarts. Redis cache settings can be fine-tuned with the QSV_REDIS_CONNSTR, QSV_REDIS_TTL_SECONDS, QSV_REDIS_TTL_REFRESH and QSV_FP_REDIS_CONNSTR environment variables.
  • Alternatively, the fetch and fetchpost commands can use a local disk cache with the --disk-cache option. The default cache directory is ~/.qsv-cache/fetch. The cache directory can be changed with the QSV_CACHE_DIR environment variable or the --disk-cache-dir command-line option. A disk cache entry is expired after 28 days by default, but this can be changed with the QSV_DISK_CACHE_TTL_SECONDS and QSV_DISKCACHE_TTL_REFRESH environment variables.
  • The luau command caches lookup tables on disk using the QSV_CACHE_DIR environment variable and the --cache-dir command-line option. The default cache directory is ~/.qsv-cache. The QSV_CACHE_DIR environment variable overrides the --cache-dir command-line option.

SIMD-accelerated UTF-8 Validation for Performance

Rust strings are utf-8 encoded. As a result, qsv REQUIRES UTF-8 encoded files.

Still, users will attempt to use non UTF-8 encoded files, and for the most part, they will still work! This is because most qsv commands use ByteRecords, where qsv manipulates raw bytes and doesn't care about the encoding.

Where it does matter, qsv will attempt to convert the bytes to UTF-8. But instead of using std::str::from_utf8, it makes extensive use of simdutf8 for SIMD-accelerated UTF-8 validation, which is up to 23x faster on x86-64 and 11x faster on aarch64 (Apple Silicon).

As UTF-8 is the de facto encoding standard, this shouldn't be a problem most of the time. However, should you need to process a CSV file with a different encoding, use the input command with the --output option first to "loosely transcode" it to UTF-8.

Nightly Release Builds

Pre-built binaries compiled using Rust Nightly/Unstable are also available for download. These binaries are optimized for size and speed:

  • compiled with the last known Rust nightly/unstable that passes all 1,400+ tests.
  • stdlib is compiled from source, instead of using the pre-built stdlib. This ensures stdlib is compiled with all of qsv's release settings (link time optimization, opt-level, codegen-units, panic=abort, etc.), presenting more opportunities for Rust/LLVM to optimize the generated code. This is why we only have nightly release builds for select platforms (the platform of GitHub's action runners), as we need access to the "native hardware" and cannot cross-compile stdlib to other platforms.
  • set panic=abort - removing panic-handling/formatting and backtrace code, making for smaller binaries.
  • enables unstable/nightly features in the rand, regex, hashbrown, pyo3 and polars crates, that unlock performance/SIMD features on those crates.

Despite the 'unstable' label, these binaries are actually quite stable, given how Rust is made and are really more about performance (that's why we can still compile with Rust stable). You only really loose the backtrace messages when qsv panics.

If you need to maximize performance - use the nightly builds. If you prefer a "safer", rock-solid experience, use the stable builds.

If you want to really squeeze every little bit of performance from qsv, build it locally like how the Nightly Release Builds are built, with the additional step of optimizing the build to your machine's CPU by setting RUSTFLAGS='-C target-cpu=native'.

Doing so will ensure CPU features are tailored to your hardware and you're using the latest Rust nightly.

For example, on Ubuntu 22.04 LTS Linux:

rustup default nightly
rustup update
export RUSTFLAGS='-C target-cpu=native'

# to build qsv on nightly with all features. The binary will be in the target/release-nightly folder.
cargo build --profile release-nightly --bin qsv -Z build-std=std,panic_abort \
  -Z build-std-features=panic_immediate_abort \
  --features all_features,nightly --target x86_64-unknown-linux-gnu

# to build qsvlite
cargo build --profile release-nightly --bin qsvlite -Z build-std=std,panic_abort \
  -Z build-std-features=panic_immediate_abort \
  --features lite,nightly --target x86_64-unknown-linux-gnu

# to build qsvdp
cargo build --profile release-nightly --bin qsvdp -Z build-std=std,panic_abort \
  -Z build-std-features=panic_immediate_abort \
  --features datapusher_plus,nightly --target x86_64-unknown-linux-gnu

With that said, there are times that Rust Nightly/Unstable does "break" qsv. That's why we include qsv_rust_version_info.txt in the nightly release build zip files, should you need to pin Rust to a specific nightly version when building locally.

Benchmarking for Performance

Use and fine-tune the benchmark script when tweaking qsv's performance to your environment. Don't be afraid to change the benchmark data and the qsv commands to something that is more representative of your workloads.

Use the generated benchmark TSV files to meter and compare performance across platforms. You'd be surprised how performance varies across environments - e.g. qsv's join performs abysmally on Windows's WSL running Ubuntu 20.04 LTS, taking 172.44 seconds. On the same machine, running in a VirtualBox VM at that with the same Ubuntu version, join was done in 1.34 seconds - two orders of magnitude faster!

However, stats performs two times faster on WSL vs the VirtualBox VM - 2.80 seconds vs 5.33 seconds for the stats_index benchmark.