4 ms·
The author's scp at the end of the article hits 10 MB/s, which is about what I cap out at when transferring over SSH from my MacBook. That's because it maxes ou
by Bognar 3y ago
The author's scp at the end of the article hits 10 MB/s, which is about what I cap out at when transferring over SSH from my MacBook. That's because it maxes out a single core just trying to encrypt all the data without hardware acceleration. Straight FTP or even HTTPS upload (which is hardware accelerated) should be even faster if they have more bandwidth.
- wahern 3y agoUnless your MacBook is 20 years old, something else is going on. Are you perhaps on 100Mb/s link? With overhead 10MB/s is about what one would expect on a 100Mb/s link. Even without hardware acceleration my 2018 Mac Mini, which is much slower than my 2020 MBA using ARM, does much better than that--64MB/s, but that's a floor as the file itself is only 64MB. I tested by forcing chacha20-poly1305, which wouldn't be 6x faster than software-only aes128-gcm or aes256-gcm. Also, though /usr/bin/ssh on macOS is compiled with libressl, and there was a brief post-fork period when libressl lacked accelerated cipher implementations, it has supported AES-NI for quite awhile. I'm sensitive to these things as I still run a few PC Engines APUs and force aes256-gcm in sshd_config; by default OpenSSH prefers chacha20-poly1305. To an APU I get 30MB/s with AES-NI aes256-gcm, and 10MB/s with chacha20-poly1305. The APU has an AMD GX-412TC, which came out in 2013 but as a low-powered embedded part has the performance of something 5 or 10 years older.
- Bognar 3y agoI misspoke, not a MacBook but a MacMini. It's on a wired 1G connection transferring across local LAN. I'll have to rerun the experiment when I'm at work again and investigate, but iirc when we did this previously we were seeing one maxed core and low scp perf.
- mytailorisrich 3y agoNot sure I follow why SSH is not hardware accelerated but HTTPS is? Aren't both using AES?