Performance considerations
The performance of the backup system is important. During backup, all backup tasks need to be performed inside of the backup window. The restore performance is important to meet service level agreements.
Compression settings
Bareos supports multiple compression algorithms. When configured, Bareos will compress the backed up data in the Bareos File Daemon before sending it over to the Bareos Storage Daemon during backup, and will decompress the data on restore before writing it to the disk.
Bareos Version >= 23.0.0, supports deploying the compression work over multiple threads.
Lab Setup
To analyze the impact of the multithreading on different compression settings and different type of data being backed up, a lab environment was set up.
Three different data sets were backed up:
Multimedia Files
A Linux root Filesystem.
The Linux Kernel 6.6 Sources
Multimedia Files |
Linux Root Filesystem |
Linux Kernel Sources |
|
|---|---|---|---|
Size |
6.4G |
6.2G |
1.5G |
Files |
295 |
149 118 |
87 084 |
Compressibility |
Bad |
Intermediate |
Good |
OS |
Fedora 39 x86_64 |
CPU |
AMD Ryzen 7 PRO 7840U w/ Radeon 780M Graphics, 16 cores |
Memory |
64 GB |
Storage Device |
2TB SSD SAMSUNG MZVL22T0HDLB nvme |
Source Filesystem |
xfs on ssd partition |
Storage Filesystem |
tmpfs 10G ramdisk |
The test setup is configured so that the different Datasets are read from a xfs partition on the nvme Disk. The backup volumes are written into a 10GB tmpfs ramdisk.
Note
The value “0” for the number of workers means the original code path used in bareos 22. Other numbers are the value of MaximumWorkersPerJob in the client configuration.
Multimedia Files
Backup
MaximumWorkersPerJob vs compression algorithms for the multimedia data backup
The test shows, that the system is capable of 650 M/s backup speed when backing up the multimedia data. The lzo compression performs best, but does not compress the data at all. The lz4 compressor reaches only 1% of compression on that data at about 150M/s
lz4 scales well with the number of workers and achieves a maximum speed of 500M/s with 5 workers. The other compression algorithms can achieve higher compression rates with much slower backup speed compared to no compression. The impact of multithreading is high on all other algorithms and can speedup the backup by the factor 5.
Overall, the backup speed is very slow compared to the uncompressed backup.
Restore
MaximumWorkersPerJob vs compression algorithms for the multimedia data restore
The restore speed is not influenced by the number of threads and reaches almost the uncompressed rate with lzo, lz4 and lz4hc. The gzip family of algorithms slows down the restore speed by around 50%.
Linux Filesystem
Backup
MaximumWorkersPerJob vs compression algorithms for the linux root fs data backup
The maximum backup speed of the Linux Filesystem is about 260MB/s. Lzo reaches a compression rate of 46.7% while slowing down the backup speed to about 50%. Using multiple threads increases the lzo speed form 127MB/s to 150MB/s (about 18%)
lz4 only reaches half of the speed of lzo while even reaching a worse compression! gzip1 is faster than lz4hc while reaching a better compression rate (56.9% vs 52.3%.)
Restore
MaximumWorkersPerJob vs compression algorithms for the linux root fs data restore
On restore of the linux rootfs, lzo, lz4 and lzfast are only a bit slower than the uncompressed backup, while the gzip algorithms are around 20% slower.
Linux Kernel Sources
Backup
MaximumWorkersPerJob vs compression algorithms for the linux kernel source data backup
The maximum backup speed of the kernel sources is 100M/s. lzo reaches a speed of 25MB/s with a compression of 68.8%. lz4 reaches about 20MB/s with a compression rate of 67.4% and so it is slower than lzo and worse in compression.
The number of threads almost has no impact on the backup speed for all algorithms.
Restore
MaximumWorkersPerJob vs compression algorithms for the linux kernel source data restore
Restoring the linux kernel sources reaches around 125MB/s for uncompressed backups. All algorithms reach about 110-120MB/s. This shows that the decompression of the data probably is not the bottleneck when restoring these files.
Conclusion
Depending on the properties of the data being backed up, the backup speed can vary. It is interesting, that the lzo algorithm shows the best performance in our tests. When the data is not compressible, it is advisable not to enable compression at all.
The results on this test system suggest that it makes sense to use lzo compression in most cases.
In case the lzo algorithm shows bad compression ratio on your data, it might make sense to disable compression at all. However,in lab environment, the bandwidth between Bareos File Daemon and Bareos Storage Daemon was not a limiting factor. In real world environments, the bandwith probably will the bottleneck and thus using compression will lead to an overall much better performance.
Regarding restores it can be said that the speed is not reduced by the lzo and lz4 family of compression algorithms, while gzip does have a significant negative impact on restore speed.
Because of their bad performance, gzip algorithms should not be used at all.
TLS Settings
Bareos encrypts all network traffic by default. TLS-PSK is used to encrypt the traffic. It is however possible to disable the TLS encryption and allow cleartext communication between.
Disabling TLS Encryption
We tested the influence of disabling TLS on the backup speed of our lab environment. Only LZO and No Compression are regarded.
No TLS with no Encryption
Speed comparison [MiB/s] when TLS is disabled and no compression is configured.
No TLS with LZO Encryption
Speed comparison [MiB/s] when TLS is disabled and LZO compression is configured.
The impact of disabling TLS depends on the dataset being backed up. The faster the backup itself runs(and the faster the network traffic itself is), the bigger is the impact. Both with cleartext and lzo encryption we see similar accelerations:
Backup |
Backup LZO |
Restore |
Restore LZO |
|
|---|---|---|---|---|
Multimedia files |
67% |
67% |
43% |
38% |
Linux Root Filesystem |
15% |
28% |
26% |
40% |
Kernel Source |
8% |
8% |
10% |
38% |
Conclusion
Disabling the TLS encryption of the network traffic in bareos can bring a significant acceleration.
The acceleration depends on the dataset being backed up. The faster the network transfer is, the more influence the TLS load has.
If the backup speed is not fast enough, and traffic encryption is not required, disabling TLS can help to get the backups faster.
Use different TLS Ciphers
If disabling TLS is not an option, it still is possible to use a different cipher that might be faster than what is chosen by default.
Use kTLS
Newer Linux and FreeBSD versions have the possibility to use Kernel TLS. Kernel TLS allows the cryptographic library to ask the Kernel to do the encryption and decryption of the network traffic. If appropriate hardware(e.g. Network Card) is available, the TLS load can be offloaded to the NIC.
Extended Attribute and ACL Support
Bareos is capable of backing up extended attributes, and the support for both is enabled by default. If the backup speed needs to be accelerated, and not backing up extended attributes and/or ACLs, both options can be disabled which will save some time for each file being backed up.
Disable ACL and Xattr Support
Speed comparison when extended attribute and acl support is disabled. Speed is in MB/s.
Backup |
Restore |
|
|---|---|---|
Multimedia files |
0% |
0% |
Linux Root Filesystem |
19% |
37% |
Kernel Source |
17% |
0% |
Conclusion
Disabling the extended attribute and acl support in the fileset can improve the backup speed. As always, it depends on the dataset being backed up, and if the acl and/or extended attribute information is expendable.
File Checksums
If the FileSet option signature is set, Bareos calculates a checksum for every file being backed up.
The following graphs show the results of testing the performance of the available checksums.
Checksum performance comparison for multimedia file dataset
Checksum performance comparison for rootfs dataset
Checksum performance comparison for kernel source dataset
Conclusion
The results of the tests are unfortunately not clear. SHA512 seems to be the fastest algorithm, though the difference between the different algorithms is not significant.
It is also strange that the restore results seem to be much less affected than the backup results, as the restore code also recalculates the checksum to verify that the restore successfully restored the correct data.
Theoretically, xxh128 is the by far fastest algorithm offered and thus should be the fastest option. The results show that more investigation is required to understand the results.
