Lab 07 · Performance

Compression is the costliest step of the backup; encryption barely registers

Date: Oct 15, 2026Server: test server, Xeon E5-2698 v4, 18 coresReading time: 2 minTopic: Performance
The question

A backup that eats the processor slows the sites down while it runs, and the customer notices. Of the three steps the engine runs on every chunk of a file (fingerprint, compression and encryption), which weighs most on the processor? Is any algorithm worth switching?

Result
8 msper MB for compression, versus 2.4 ms for the hash and 1.1 ms for encryption, on one core
~17%more throughput if the hash becomes BLAKE3

Why it matters to whoever runs the server

  • Knowing which step weighs most shows where to save: compression costs 8 ms per MB on one core; encryption, 1.1 ms. Encrypting the backup costs the server almost nothing extra.
  • The algorithm is a trade between processor and storage space: s2 is 8× faster but takes about 27% more space. Salvo keeps light compression as the default.
Cost of each step on one core, with the defaultsMilliseconds per MB · lower is better
Compression (zstd-fastest)
8.1 ms
Hash (BLAKE2B)
2.4 ms
Encryption (AES-256-GCM)
1.1 ms
Source: Salvo Lab · Salvo engine benchmark, processor only · Oct 15, 2026
See the data as a table
Algorithm1 coreAll cores
Hash BLAKE31.5 GB/s18.7 GB/s
Hash BLAKE2B (default)418 MB/s5.7 GB/s
Encryption AES-256-GCM (default)900 MB/s7.7 GB/s
Encryption ChaCha20-Poly1305654 MB/s7.4 GB/s
Compression zstd-fastest (default), 15.4% of the size123 MB/s848 MB/s
Compression zstd, 13% of the size89 MB/s650 MB/s
Compression s2, 19.6% of the size1 GB/s2.9 GB/s

Method

Test server with a Xeon E5-2698 v4 (2.2 GHz, 18 cores).

The Salvo engine’s benchmark tool, measuring only the processor, with no disk and no network.

Compression was measured on 256 MB of real code from a WordPress site.

What we decided

We keep AES-256 and zstd-fastest. Before switching the default hash to BLAKE3, we’ll measure it on a new repository: the hash is fixed when the repository is created. s2 is 8× faster but takes ~27% more space; it may become an option for those who prefer speed over space.

?
What we don’t know yetThis is a pure processor measurement: in a real backup, reading the disk and walking thousands of files also weighs in. The sample is only site code, mostly text; whole accounts, with images and email, compress less. And single-core numbers vary with server load (BLAKE3 gave 0.9 to 1.5 GB/s across two runs); the ranking between algorithms held.
Want to see Salvo on your server?Join the beta waitlist. We’ll let you know when there’s a spot.

How it connects to the other labs

Faster backups
30 MB30 MB of cache for 233 thousand files, and deleting it cost nothingLab 01 · Oct 14, 20263×The limit was the encryption, not the network: switching ciphers tripled the uploadLab 06 · Oct 15, 2026141The engine uploads 141 MB/s reading one file per core; 4 accounts together used 2.4 GB of memoryLab 08 · Oct 15, 202633 s3 full cPanel accounts in 33 s on a production server; the incremental, in 7 sLab 09 · Oct 15, 20264 sThe next-day backup, before dawn and cold: 4 s for 21 thousand filesLab 12 · Oct 16, 20267.7×The next-day backup on the same 6 accounts: 494 s on JetBackup, 64 s on SalvoLab 13 · Oct 16, 2026166×The first backup side by side: 45% less time and 166× fewer files on the storageLab 14 · Oct 16, 202699.98%WordPress backup with email: 6 real sites restored by Salvo and by JetBackupLab 23 · Oct 19, 2026
salvo. Backup for hosting servers. A GK2 product.
Product
Lab
Salvo
Language
© 2026 GK2We use analytics cookies (Google Analytics).salvobackup.com/lab/compression-is-the-costliest-step/