htcat is a utility to perform parallel, pipelined execution of a
single HTTP GET. htcat is intended for the purpose of
incantations like:
htcat https://host.net/file.tar.gz | tar -zx
It is tuned (and only really useful) for faster interconnects:
$ htcat http://test.com/file | pv -a > /dev/null
[ 109MB/s]
This is on a gigabit network, between an AWS EC2 instance and S3. This represents 91% use of the theoretical maximum of gigabit (119.2 MiB/s).
This program depends on Go 1.1 or later. One can use go get to
download and compile it from source:
$ go get github.com/htcat/htcat/cmd/htcat
For correspondence of all sorts, write to [email protected]. Bugs can be filed at htcat's GitHub Issues page.
htcat works by determining the size of the Content-Length of the
URL passed, and then partitioning the work into a series of GETs
that use the Range header in the request, with the notable exception
of the first issued GET, which has no Range header and is used to
both start the transfer and attempt to determine the size of the URL.
Unlike most programs that do similar Range-based splitting, the
requests that are performed in parallel are limited to some bytes
ahead of the data emitted so far instead of splitting the entire byte
stream evenly. The purpose of this is to emit those bytes as soon as
reasonably possible, so that pipelined execution of another tool can,
too, proceed in parallel.
These requests may complete slightly out of order, and are held in reserve until contiguous bytes can be emitted by a defragmentation routine, that catenates together the complete, consecutive payloads in memory for emission.
Tweaking the number of simultaneous transfers and the size of each
GET makes a trade-off between latency to fill the output pipeline,
memory usage, and churn in requests and connections and incurring
their associated start-up costs.
If htcat's peer on the server side processes Range requests more
slowly than regular GET without a Range header, then, htcat's
performance can suffer relative to a simpler, single-stream GET.
These are measurements falling well short of real benchmarks that are intended to give a rough sense of the performance improvements that may be useful to you. These were taken via an AWS EC2 instance connecting to S3, and there is definitely some variation in runs, sometimes very significant, especially at the higher speeds.
| Tool | TLS | Rate |
|---|---|---|
| htcat | no | 109 MB/s |
| curl | no | 36 MB/s |
| aria2c -x5 | no | 113 MB/s |
| htcat | yes | 59 MB/s |
| curl | yes | 5 MB/s |
| aria2c -x5 | yes | 17 MB/s |
On somewhat small files, the situation changes: htcat chooses
smaller parts, as to still get some parallelism.
Below are results while performing a 13MB transfer from S3 (Seattle) to an EC2 instance in Virginia. Notably, TLS being on or off did not seem to matter, perhaps in this case it was not a bottleneck.
| Tool | Time |
|---|---|
| curl | 5.20s |
| curl | 7.75s |
| curl | 6.36s |
| htcat | 2.69s |
| htcat | 2.50s |
| htcat | 3.25s |
Results while performing a transfer of the same 13MB file from S3 to EC2, but all within Virginia:
| Tool | TLS | Time |
|---|---|---|
| curl | no | 0.29s |
| curl | no | 0.75s |
| curl | no | 0.44s |
| htcat | no | 0.30s |
| htcat | no | 0.30s |
| htcat | no | 0.48s |
| curl | yes | 2.69s |
| curl | yes | 2.69s |
| curl | yes | 2.62s |
| htcat | yes | 1.37s |
| htcat | yes | 0.45s |
| htcat | yes | 0.59s |
Results while performing a 4.6MB transfer on a fast (same-region)
link. This file is small enough that htcat disables multi-request
parallelism. Given that, it's unclear why htcat performs markedly
better on the TLS tests than curl.
| Tool | TLS | Time |
|---|---|---|
| curl | no | 0.14s |
| curl | no | 0.13s |
| curl | no | 0.14s |
| htcat | no | 0.23s |
| htcat | no | 0.16s |
| htcat | no | 0.17s |
| curl | yes | 0.95s |
| curl | yes | 0.97s |
| curl | yes | 0.99s |
| htcat | yes | 0.38s |
| htcat | yes | 0.34s |
| htcat | yes | 0.24s |