Performance is slower than sshfs

[LionsAd]

Great project - looks really nice.

I would have expected 9p with the nice rust tokio runtime to easily beat sshfs, but it turns out sshfs is faster by factor of 3x.

Note: I am using a Mac and a qemu VM as that is the use case for me.

Anything I can tweak in terms of using more CPUs and better threading for the tokio co-routines?

Edit: Here are the benchmarks (all done on QEMU on Silicon Mac):

TL;DR:

• SSHFS has the lowest latency and okay raw data throughput
• virtio has the best throughput for raw data writing by far
• rust-p9 is unfortunately slowest right now, latency for find is comparable with virtio so this might be a 9P protocol limitation

Setup

QEMU option for virtio:

-virtfs local,path=projects/oss/rust-9p,mount_tag=foo,id=foo,security_model=mapped-xattr

All 9p mounts have been done with:

# virtio
sudo mount -t 9p -o version=9p2000.L,trans=virtio,uname=$USER,msize=104857600,cache=mmap foo /mnt

# rust-9p
sudo mount -t 9p -o version=9p2000.L,trans=tcp,port=8564,uname=$USER,msize=104857600,cache=mmap 192.168.5.2 /mnt

and the target directory is the one of rust-9p.

Benchmarks

9p-virtio:

colima:~$ sh /mnt/perf.sh 
+ cd /mnt/foo
+ time find target
real	0m 0.78s
user	0m 0.00s
sys	0m 0.10s
+ time cp -r target target3
real	0m 8.83s
user	0m 0.00s
sys	0m 1.10s
+ time rm -rf target3
real	0m 1.56s
user	0m 0.00s
sys	0m 0.30s
+ time dd 'if=/dev/zero' 'of=test.dat' 'bs=1G' 'count=3'
3+0 records in
3+0 records out
real	0m 1.72s
user	0m 0.00s
sys	0m 0.70s

SSHFS:

+ cd /mnt/foo
+ time find target
real	0m 0.45s
user	0m 0.04s
sys	0m 0.00s
+ time cp -r target target3
real	0m 5.28s
user	0m 0.03s
sys	0m 0.62s
+ time rm -rf target3
real	0m 0.74s
user	0m 0.00s
sys	0m 0.09s
+ time dd 'if=/dev/zero' 'of=test.dat' 'bs=1G' 'count=3'
3+0 records in
3+0 records out
real	0m 17.55s
user	0m 0.00s
sys	0m 2.26s

https://github.com/pfpacket/rust-9p via TCP transport:

colima:/$ sh /mnt/perf.sh 
+ cd /mnt/foo
+ time find target
real	0m 0.77s
user	0m 0.00s
sys	0m 0.11s
+ time cp -r target target3
real	0m 10.25s
user	0m 0.00s
sys	0m 1.45s
+ time rm -rf target3
real	0m 1.40s
user	0m 0.03s
sys	0m 0.17s
+ time dd 'if=/dev/zero' 'of=test.dat' 'bs=1G' 'count=3'
3+0 records in
3+0 records out
real	1m 8.48s
user	0m 0.00s
sys	0m 5.01s
colima:/$ 

[LionsAd]

Added benchmarks ^^

[ERnsTL]

Greetings everyone, I also noticed this like you @LionsAd - tried doing "ls" of a remote directory in cable LAN shared via unpfs and Linux (Ubuntu 22.04) with the exact mount settings given in the Readme file, latest rustc via rustup etc. - compared to sshfs there is a noticeable delay. This is unexpected.