debian-mirror-gitlab/doc/administration/operations/filesystem_benchmarking.md
2022-11-25 23:54:43 +05:30

128 lines
4.5 KiB
Markdown

---
stage: Systems
group: Distribution
info: To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/product/ux/technical-writing/#assignments
---
# File system performance benchmarking **(FREE SELF)**
File system performance has a big impact on overall GitLab performance,
especially for actions that read or write to Git repositories. This information
helps benchmark file system performance against known good and bad real-world
systems.
Normally when talking about file system performance the biggest concern is
with Network File Systems (NFS). However, even some local disks can have slow
I/O. The information on this page can be used for either scenario.
## Executing benchmarks
### Benchmarking with `fio`
We recommend using
[Fio](https://fio.readthedocs.io/en/latest/fio_doc.html) to test I/O
performance. This test should be run both on the NFS server and on the
application nodes that talk to the NFS server.
To install:
- On Ubuntu: `apt install fio`.
- On `yum`-managed environments: `yum install fio`.
Then run the following:
```shell
fio --randrepeat=1 --ioengine=libaio --direct=1 --gtod_reduce=1 --name=test --bs=4k --iodepth=64 --readwrite=randrw --rwmixread=75 --size=4G --filename=/path/to/git-data/testfile
```
This creates a 4GB file in `/path/to/git-data/testfile`. It performs
4KB reads and writes using a 75%/25% split in the file, with 64
operations running at a time. Be sure to delete the file after the test
completes.
The output varies depending on what version of `fio` installed. The following
is an example output from `fio` v2.2.10 on a networked solid-state drive (SSD):
```plaintext
test: (g=0): rw=randrw, bs=4K-4K/4K-4K/4K-4K, ioengine=libaio, iodepth=64
fio-2.2.10
Starting 1 process
test: Laying out IO file(s) (1 file(s) / 1024MB)
Jobs: 1 (f=1): [m(1)] [100.0% done] [131.4MB/44868KB/0KB /s] [33.7K/11.3K/0 iops] [eta 00m:00s]
test: (groupid=0, jobs=1): err= 0: pid=10287: Sat Feb 2 17:40:10 2019
read : io=784996KB, bw=133662KB/s, iops=33415, runt= 5873msec
write: io=263580KB, bw=44880KB/s, iops=11219, runt= 5873msec
cpu : usr=6.56%, sys=23.11%, ctx=266267, majf=0, minf=8
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=0.1%, 32=0.1%, >=64=100.0%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.1%, >=64=0.0%
issued : total=r=196249/w=65895/d=0, short=r=0/w=0/d=0, drop=r=0/w=0/d=0
latency : target=0, window=0, percentile=100.00%, depth=64
Run status group 0 (all jobs):
READ: io=784996KB, aggrb=133661KB/s, minb=133661KB/s, maxb=133661KB/s, mint=5873msec, maxt=5873msec
WRITE: io=263580KB, aggrb=44879KB/s, minb=44879KB/s, maxb=44879KB/s, mint=5873msec, maxt=5873msec
```
Notice the `iops` values in this output. In this example, the SSD
performed 33,415 read operations per second and 11,219 write operations
per second. A spinning disk might yield 2,000 and 700 read and write
operations per second.
### Simple benchmarking
NOTE:
This test is naive but can be used if `fio` is not
available on the system. It's possible to receive good results on this
test but still have poor performance due to read speed and various other
factors.
The following one-line commands provide a quick benchmark for file system write and read
performance. This writes 1,000 small files to the directory in which it is
executed, and then reads the same 1,000 files.
1. Change into the root of the appropriate
[repository storage path](../repository_storage_paths.md).
1. Create a temporary directory for the test so it can be removed later:
```shell
mkdir test; cd test
```
1. Run the command:
```shell
time for i in {0..1000}; do echo 'test' > "test${i}.txt"; done
```
1. To benchmark read performance, run the command:
```shell
time for i in {0..1000}; do cat "test${i}.txt" > /dev/null; done
```
1. Remove the test files:
```shell
cd ../; rm -rf test
```
The output of the `time for ...` commands resemble the following. The
important metric is the `real` time.
```shell
$ time for i in {0..1000}; do echo 'test' > "test${i}.txt"; done
real 0m0.116s
user 0m0.025s
sys 0m0.091s
$ time for i in {0..1000}; do cat "test${i}.txt" > /dev/null; done
real 0m3.118s
user 0m1.267s
sys 0m1.663s
```
From experience with multiple customers, this task should take under 10
seconds to indicate good file system performance.