tahoe-lafs/trac-2024-07-25

automated comparative 'k' performance tests and graphs #1530

New issue

Open

opened 2011-09-07 16:46:19 +00:00 by warner · 5 comments

warner commented

2011-09-07 16:46:19 +00:00

Owner

Here's my plan for some automated performance tests I intend to build over the next few months:

use the four servers on loan from Atlas Networks, one as a client, three as storage servers
use the attached upload-kN.py to upload a bunch of files with various sizes and encoding parameters (actually just varying 'k' and leaving N/k constant, since I don't think N affects performance much and would increase the runtime drastically).
write a script which fetches the directory contents once to get all the filecaps, then randomly downloads files from it, keeping track of how long each one takes (also recording other download-result timing data)
push that data to a CouchDB server for long-term storage (probably on iriscouch.com because it's free for small datasets, but it's easy to move a CouchDB to a different server later)
write some static HTML, served by that same CouchDB server, that uses d3.js to construct graphs. The script in the HTML will turn around and fetch the relevant data from the DB with XHR.
make a scatterplot of time vs k, then a candlestick chart

Later, I hope to use the same approach to store+display performance data across revisions, like the speedtest and memtest data we're getting from the buildbot.

Here's my plan for some automated performance tests I intend to build over the next few months: * use the four servers on loan from Atlas Networks, one as a client, three as storage servers * use the attached [upload-kN.py](/tahoe-lafs/trac-2024-07-25/attachments/000078ac-0ad5-4e42-9b2b-531b8128d5b7) to upload a bunch of files with various sizes and encoding parameters (actually just varying 'k' and leaving N/k constant, since I don't think N affects performance much and would increase the runtime drastically). * write a script which fetches the directory contents once to get all the filecaps, then randomly downloads files from it, keeping track of how long each one takes (also recording other download-result timing data) * push that data to a CouchDB server for long-term storage (probably on iriscouch.com because it's free for small datasets, but it's easy to move a CouchDB to a different server later) * write some static HTML, served by that same CouchDB server, that uses [d3.js](http://mbostock.github.com/d3/) to construct graphs. The script in the HTML will turn around and fetch the relevant data from the DB with XHR. * make a scatterplot of time vs k, then a candlestick chart Later, I hope to use the same approach to store+display performance data across revisions, like the speedtest and memtest data we're getting from the buildbot.