Performance benchmarking

DISK performance

• IOPS : I/O operations per second
• Throughput or Bandwidth : the amount of data transferred per second
• Latency  : the time required to perform an I/O operation

hdparm

hdparm is a tool for collecting and manipulating the low-level parameters of SATA/PATA/SAS and older IDE storage devices. It also contains a benchmark mode we can invoke with the -t flag, even on a disk with mounted partitions:

$ sudo hdparm -t /dev/nvme0n1
/dev/nvme0n1:
Timing buffered disk reads: 10014 MB in  3.00 seconds = 3337.66 MB/sec
 
*** the disk read output means how fast the disk can sustain sequential data reads without any filesystem overhead and without prior caching since hdparm automatically flushes it

Using dd for disk performance tests is not recommended. Because dd will perform a sequential write what is not a real scenario, To evaluate your servers, use fio or bonnie++ for performance tests.

sysbench

## step 1, prepare benchmark dir and files
## –file-total-size flag indicates the total size of test files to be created locally. 
$mkdir temp
$ cd temp
$ sysbench fileio --file-total-size=50G prepare

## step 2, Execute benchmark 
sysbench --file-total-size=50G --file-test-mode=rndrw --file-extra-flags=direct fileio run

–file-extra-flags=direct option tells sysbench to use direct I/O, which will bypass the page cache. This choice ensures that the test interacts with the disk and not with the main memory
–file-test-mode accepts the following workloads:
    seqwr → sequential write
    seqrewr → sequential rewrite
    seqrd → sequential read
    rndrd → random read
    rndwr → random write
    rndrw → combined random read/write

## step 3, clean up
$ sysbench fileio cleanup

fio

fio stands for Flexible I/O Tester and refers to a tool for measuring I/O performance. With fio, we can set our workload to get the type of measurement we want.

$sudo fio --filename=/dev/sda --readonly --rw=read --name=TEST --runtime=3

–filename flag allows us to test a block device, such as /dev/sda, or just a file
–runtime=3 means three seconds
–name : test name
-rw options are 
    read → sequential reads
    write → sequential writes
    randread → random reads
    randwrite → random writes
    rw, readwrite → mixed sequential reads and writes
    randrw → mixed random reads and writes

fio test can be done through fio file, for example

$cat test.fio 
[global]
runtime=3
name=TEST
rw=read
	
[job1]
filename=/dev/nvme0n1

$ fio --readonly test.fio

iozone test

iozone -+m <conf_file> -i 0 -w -+n -c -C -e -s 4g -r 1024k -t 8

Actual results:

gluster-block:
        Children see throughput for  8 initial writers  =   87671.01 kB/sec
        Children see throughput for  8 readers          =  133351.47 kB/sec

gluster-fuse:
        Children see throughput for  8 initial writers  =  939760.11 kB/sec
        Children see throughput for  8 readers          =  791956.57 kB/sec

gluster-nfs:
        Children see throughput for  8 initial writers  =  989822.15 kB/sec
        Children see throughput for  8 readers          = 1203338.91 kB/sec

PCI Switch, CPU and GPU Direct server topology

Reference

<reference/>