Poor performance copying to AFP share with large number of files

[Kevin Horton]

I'm troubleshooting poor performance copying from macOS 10.12.2 to an AFP share on a box running FreeNAS-9.10.1-U4, with 1GbE networking. The AFP share in question has quite a large number of files and directories. This post suggested that AFP shares with smaller numbers of files and directories had much better performance, and that does seem to be true. A new share to an empty dataset averages around 800 Mb/s when transferring a large file, whereas copying to the share with more files has very variable performance that bounces around between 100 and 400 Mb/s.

I'm wondering if the poor performance with large AFP shares is a problem with macOS, or with Netatalk on FreeNAS. Can anyone suggest any tweaks I could make either on FreeNAS or macOS to improve performance? For the moment, if I have a large amount of stuff to copy to FreeNAS, I'll send it to my small share, or use rsync. I'll also do some testing with CIFS. I had tested it when I first set up FreeNAS, and it performed worse than AFP at that time, but that was with FreeNAS 9.3, OS X 10.10, and much smaller shares. A lot has changed since then.

 

[Ericloewe]

Have you checked local storage performance with small files? That may well be your bottleneck.

 

[balamut]

Megabits or megabytes?

 

[Kevin Horton]

Have you checked local storage performance with small files? That may well be your bottleneck.

I'm testing with a 29 GB tar of a VM. I fully expect much worse performance with large numbers of small files, which is why I tar'd the VM before backing it up.

Megabits or megabytes?

Megabits. If I was getting up to 800 Megabytes/second on GbE that would be rather incredible.

I'm getting 100 to 400 megabits/second copying to the large AFP share, and about 800 megabits/second copying to the empty share. I'm satisfied with the performance going to the small share, as it is close enough to GbE performance, considering various overheads. I'm trying to figure out why things are so much worse when copying to the large share.

 

[Kevin Horton]

I just tried with a CIFS share. I deleted the AFP share for the large dataset, and created a CIFS share to the same dataset. It hit a solid 400 megabits/second. It was like it hit some ceiling, as the performance was very steady. Better than AFP to the same dataset, but I don't understand why it maxed out at 400 Mb/s. testparm shows "max protocol = SMB2", if that makes a difference.