Tutorial

Iozone runs

You will need to create iozone result files first. We recommend to perform 5 runs to create one set of data.

Baseline set of iozone data

iozone -U /mnt/test -a -f /mnt/test/iozone -n 4k -g 4096m > baseline_1.iozone
iozone -U /mnt/test -a -f /mnt/test/iozone -n 4k -g 4096m > baseline_2.iozone
iozone -U /mnt/test -a -f /mnt/test/iozone -n 4k -g 4096m > baseline_3.iozone
iozone -U /mnt/test -a -f /mnt/test/iozone -n 4k -g 4096m > baseline_4.iozone
iozone -U /mnt/test -a -f /mnt/test/iozone -n 4k -g 4096m > baseline_5.iozone

Target set of iozone data

iozone -U /mnt/test -a -f /mnt/test_1/iozone -n 4k -g 4096m > target_1.iozone
iozone -U /mnt/test -a -f /mnt/test_1/iozone -n 4k -g 4096m > target_2.iozone
iozone -U /mnt/test -a -f /mnt/test_1/iozone -n 4k -g 4096m > target_3.iozone
iozone -U /mnt/test -a -f /mnt/test_1/iozone -n 4k -g 4096m > target_4.iozone
iozone -U /mnt/test -a -f /mnt/test_1/iozone -n 4k -g 4096m > target_5.iozone

Target set of iozone result files can be produced for example

• on another version of Linux kernel
• on another physical device with the same file-system as for baseline
• on the same physical device with the different file-system as for baseline. It can include for example encryption layer (LUKS).

Analyses with help of iozone_results_comparator.py

The typical usage is

1. run iozone_results_comparator.py in the default mode * Check summary graph * Check default results for all operations or interesting operations based on the overall summary graph
2. comparison of more than two data sets using the multiset mode plots

BASELINE=$(find ./ -name "baseline*.iozone" | sort)
SET1=$(find ./ -name "target*.iozone" | sort)

#Default analyses mode (overview)
~/GIT/iozone-results-comparator/src/iozone_results_comparator.py --baseline ${BASELINE} --set1 ${SET1} 

Please note you need matplotlib in version 1.2 or higher in order to get the tool working properly.

Checking default analyses mode results

Open results in your preferred HTML browser firefox html_out/index.html Check the Summary graph on the top of the page.

You will note that results in the baseline and target set differs significantly for the record rewrite operation. Let's click on the Record rewrite link in the header of the table located below the summary plot, which brings us to the Record rewrite section. There are four graphs:

Throughput as a function of the file size

In this plot there are data aggregated over all block sizes. You can see that performance starts to be different for file sizes bigger than 512kB. Median, first and third quartile values are displayed when mouse pointer is placed over the plot.

Throughput as a function of the block size

Data are aggregated over all file sizes. You can see that performance differs from block size bigger than 8kB.

Linear regression plot

In this plot each point represents throughput for one file size and one block size. X coordinate represents arithmetic average of the throughput in the baseline data set for the one particular file and block size. Y coordinate represents arithmetic average of the throughput in the target data set for the same file and block size. You can see that almost all points are above line Y=X which means that performance in the target data set is greater than performance in the baseline data set. There are however few red points where performance in target data set is slightly worse than in the baseline data set.

Percentual difference plot

File size values are displayed on the X axis and Block size values on the Y axis while the percentual difference of Baseline and Set1 is displayed in form of a circles. Bigger the difference, bigger the circle. Color of the circle represents category as shown in the legend box. Below example actually belongs to different operation to demonstrate it's function in better way. Set1 is more than 30% faster for file sizes around 2¹² bytes, but Baseline is from 15% to 30% faster for files having 2²⁰ bytes in size and also faster for files smaller than 64 bytes.

Tabbed delimited files for the further analyses

iozone_results_comparator will create also directory called tsv_out with text files containing iozone result values delimited by Tab key. These can be used for the further analyses for example in the GNU R language or in the open office.

Checking the multiset mode results

The multiset mode is enabled by using the --multiset option and providing more than two sets for comaprision.

BASELINE=$(find ./ext4 -name "*.iozone" | sort)
SET1=$(find ./xfs -name "*.iozone" | sort)
SET2=$(find ./ext2 -name "*.iozone" | sort)
SET3=$(find ./ext3 -name "*.iozone" | sort)

~/GIT/iozone-results-comparator/src/iozone_results_comparator.py --baseline ${BASELINE} --set1 ${SET1} --set2 ${SET2} --set3 ${SET3} --multiset

Let's check results in the HTML browser:

firefox html_out/index.html &

You can see throughput as a function of the file size and throughput as a function of the block size plots for every operation; every plot containing values for all the data sets as shown in the bellow examples. The plot showing throughput as a function of file size demonstrates significant performance drop for Set2 on 4G file size. If such drop needs to be analysed in more detail, users are encouraged to use the normal mode to compare the investigated data set against selected baseline.