Skip to content

Latest commit

 

History

History
170 lines (128 loc) · 7.38 KB

sampler.md

File metadata and controls

170 lines (128 loc) · 7.38 KB
Raw

Trace Sampler

Setup

# install pre-requisites
sudo apt update
sudo apt -y install git-lfs pip xz-utils
git lfs install

# clone this repo to ./sampler
cd sampler
git lfs fetch
git lfs checkout
pip install -r ../requirements.txt

Pre-processing the original trace (mandatory)

Description

The pre-processing logic works in 3 phases, first two of which perform cleaning and transforming of the original trace, whereas the third phase produces an excerpt of the clean trace, according to the user-defined start and end time.

The original trace files (invocations, durations, memory) may contain different applications and functions (i.e., their hashes). Hence, the first step is to derive a set of rows for each file with functions and applications that appear in all files. Since the original memory spec file contains rows per application and not per function, the second step is transforming the file to contain per-function rows: all memory characteristics of an application in the original trace are divided evenly among the rows, each of which corresponds to one of the application's functions.

The third preprocessing step is producing a trace excerpt with the user-defined start time and duration of the cleaned trace.

Workflow

First, download the original trace files from here and extract the CSV files (default location: data/azure/).

Then, run the following command to preprocess the trace.

python3 -m sampler preprocess -h

usage:  preprocess [-h] [-t path] -o path -s start -dur duration

optional arguments:
  -h, --help            show this help message and exit
  -t path, --trace path
                        Path to the Azure trace files
  -o path, --output path
                        Output path for the preprocessed traces
  -s start, --start start
                        Time in dd:hh:mm format, at which the excerpt of the postprocessed trace should begin, first day is day 0
  -dur duration, --duration duration
                        Duration in minutes of the excerpt extracted from the postprocessed trace

Sampling

Description

The sampling algorithm requires a clean, i.e., preprocessed, trace, which it derives samples. The user specifies the size range for samples to be generated, specifying the minimum and maximum sizes as well as the step of the sweep. The user also specifies the number of random sampling trials for each sample size. For each trial, the sample is evaluated by computing its Wasserstein distance (WD) from the original trace -- for each minute in the trace -- in two dimensions, namely the number of invocations in a minute and the amount of CPU and memory resources used by the functions. The latter is estimated (at the minute granularity for each function, which then are summed up) by multiplying the number of invocations by the average (mean) duration and by the memory footprint.

Note that the latter (resource) WD metric is normalized by 1000 000 to be in the same ballpark as the former ( invocation) WD metric. The latter metric is found empirically to be 1000 000 higher than the former because function invocation duration average is 1000ms, invocation count average is <10/min, and memory average is 200MB.

First, the sampler derives the largest sample, for which it makes several attempts (trials) when choosing the sample with the smallest aggregate WD distance (mean of the invocation and resource WDs). Then, each smaller sample is derived from the previous, larger sample; hence guaranteeing that large samples always include smaller samples that guarantees monotonic load increase (in terms of resource usage) when sweeping the sample size.

Workflow

python3 -m sampler sample -h

usage: sample [-h] -t path -orig path -o path [-min integer] [-st integer] [-max integer] [-tr integer]

options:
  -h, --help            show this help message and exit
  -t path, --source_trace path
                        Path to trace to draw samples from
  -orig path, --original_trace path
                        Path to the Azure (or other original) trace files, required to maximize the derived sample's representativity (WD from the original trace)
  -o path, --output path
                        Output path for the resulting samples
  -min integer, --min-size integer
                        Minimum sample size (#functions).
  -st integer, --step-size integer
                        Step (#functions) in sample size during sampling.
  -max integer, --max-size integer
                        Maximum sample size (#functions).
  -tr integer, --trial-num integer
                        Number of sampling trials for each sample size.

Reference traces

The reference traces are stored in data/traces/reference folder of this repository, as preprocessed_150.tar.gz and sampled_150.tar.gz files stored in Git LFS.

preprocessed_150.tar.gz contains the preprocessed traces for the original Azure trace for day 1, 09:00:00-11:30:00 (150 minutes total). 150 minutes trace captures approximately half of all functions from original Azure trace, but makes it more suitable to run in shorter experiments (10 minutes - 2 hours).

sampled_150.tar.gz contains the sampled traces for preprocessed trace from preprocessed_150.tar.gz. Sample sizes are 10-200 functions with step 10, 200-3k with step 50, and 3k-24k with step 1k.

You can untar the tarballs with the following commands:

tar -xvzf sampled_150.tar.gz
tar -xvzf preprocessed_150.tar.gz

The reference traces were obtained by running the following commands:

python3 -m sampler preprocess  -t data/azure/ -o data/traces/reference/preprocessed_150 -s 00:09:00 -dur 150

python3 -m sampler sample -t data/traces/reference/preprocessed_150 -orig data/traces/reference/preprocessed_150 -o data/traces/reference/sampled_150 -min 3000 -st 1000 -max 24000 -tr 16
python3 -m sampler sample -t data/traces/reference/sampled_150/samples/3000 -orig data/traces/reference/preprocessed_150 -o data/traces/reference/sampled_150 -min 200 -st 50 -max 3000 -tr 16
python3 -m sampler sample -t data/traces/reference/sampled_150/samples/200 -orig data/traces/reference/preprocessed_150 -o data/traces/reference/sampled_150 -min 10 -st 10 -max 200 -tr 16

Tools

Plotting

Note: Currently plotting is broken and has been disabled. (issue filled)

python3 -m sampler plot -h

usage: sampler plot [-h] -t path -k [invocation, runtime, memory] -s path

optional arguments:
  -t path, --trace path
                        Path to the trace
  -k [invocation, runtime, memory], --kind [invocation, runtime, memory]
                        Generate CDF for a single dimension
  -s path, --sample path
                        Path to the sample to be visualised
  -o path, --output path
                        Output path for the produced figures

Timeline analysis

Tools that can be used to generate the timeline of a trace are available in the tools directory.

A jupyter notebook containing examples of the timeline analysis is available in plotTimeline/analysis.ipynb