io test

io test

# we no longer use /tmp sinc it is a tmpfs / in memory

mkdir ~/tmp
cd ~/tmp

sysbench --file-fsync-freq=1 --num-threads=16 --test=fileio --file-total-size=1G --file-test-mode=rndrw --file-rw-ratio=2 cleanup > /dev/null
sysbench --file-fsync-freq=1 --num-threads=16 --test=fileio --file-total-size=1G --file-test-mode=rndrw --file-rw-ratio=2 prepare > /dev/null
sysbench --file-fsync-freq=1 --num-threads=16 --test=fileio --file-total-size=1G --file-test-mode=rndrw --file-rw-ratio=2 run

# when you are done testing
rm -fr ~/tmp

sysbench 1.0.17
Manjaro KDE

File operations:
    reads/s:                      25052.25
    writes/s:                     12526.18
    fsyncs/s:                     1603543.75

Throughput:
    read, MiB/s:                  391.44
    written, MiB/s:               195.72

General statistics:
    total time:                          10.0015s
    total number of events:              16414447

Latency (ms):
         min:                                    0.00
         avg:                                    0.01
         max:                                   31.49
         95th percentile:                        0.00
         sum:                                89075.17

Threads fairness:
    events (avg/stddev):           1025902.9375/7516.36
    execution time (avg/stddev):   5.5672/0.06

Current setup

File operations:
    reads/s:                      1200.92
    writes/s:                     600.46
    fsyncs/s:                     77052.73

Throughput:
    read, MiB/s:                  18.76
    written, MiB/s:               9.38

General statistics:
    total time:                          10.0119s
    total number of events:              787859

Latency (ms):
         min:                                    0.00
         avg:                                    0.20
         max:                                   17.89
         95th percentile:                        0.05
         sum:                               156864.48

Threads fairness:
    events (avg/stddev):           49241.1875/1430.80
    execution time (avg/stddev):   9.8040/0.01

Desktop on /tmp (Manjaro XFCE)

File operations:
    reads/s:                      31486.92
    writes/s:                     15742.96
    fsyncs/s:                     2015299.46

Throughput:
    read, MiB/s:                  491.98
    written, MiB/s:               245.98

General statistics:
    total time:                          10.0005s
    total number of events:              20625656

Latency (ms):
         min:                                    0.00
         avg:                                    0.00
         max:                                   10.70
         95th percentile:                        0.01
         sum:                                80034.78

Threads fairness:
    events (avg/stddev):           1289103.5000/5151.70
    execution time (avg/stddev):   5.0022/0.01

Desktop on /home (Manjaro XFCE)

File operations:
    reads/s:                      264.97
    writes/s:                     132.53
    fsyncs/s:                     17164.25

Throughput:
    read, MiB/s:                  4.14
    written, MiB/s:               2.07

General statistics:
    total time:                          10.0722s
    total number of events:              174849

Latency (ms):
         min:                                    0.00
         avg:                                    0.91
         max:                                   12.44
         95th percentile:                        2.43
         sum:                               159964.31

Threads fairness:
    events (avg/stddev):           10928.0625/1133.66
    execution time (avg/stddev):   9.9978/0.00

Ok the mysterie of /tmp is solved - https://wiki.manjaro.org/index.php?title=Fstab#tmpfs

mount -l | grep /tmp
tmpfs on /tmp type tmpfs (rw,noatime)

Means that doing a sysbench in /tmp is basically doing a memory i/o test, instead of doing a disk i/o test - that is why the values are so good.

That is also what finally made the differences in "performance" when we tested the different distos like ubuntu and so on, so of them do not use tmpfs for /tmp out of the box, resulting in "bad performance"