Speed issues on low-end device compared to Samba4

[bademux]

Hi,
Can we tune something for low power routers to match Samba4 in performance?
I'm using Xiaomi mir3g v1 (OpenWrt 24.10.0) with usb disk,
Ubuntu 24.10 and Windows 11 as clients.

Stock ksmdb and samba4 are comparable ~25MB/s read

For Samba 4 with I got solid 38MB/s (up to 43MB/s) read with extra configuration below:
smb.conf

aio read size = 0
aio write size = 0
use sendfile = yes

Wat can be tuned to achieve the same performance with ksmbd?

thanks.

[namjaejeon]

For Samba 4 with I got solid 38MB/s (up to 43MB/s) read with extra configuration below

What are these parameters ? These ones effect performance improvement on samba4 ?

[bademux]

Hi,
sorry I should reference the smb config.

seems like asynchronous I/O support it turns off async. It adds ~2MB/s.

aio read size = 0
aio write size = 0

Main benefit is gained with use sendfile = yes.
Documentation states "If this parameter is yes, and the sendfile() system call is supported by the underlying operating system, then some SMB read calls (mainly ReadAndX and ReadRaw) will use the more efficient sendfile system call for files that are exclusively oplocked. This may make more efficient use of the system CPU's and cause Samba to be faster. Samba automatically turns this off for clients that use protocol levels lower than NT LM 0.12 and when it detects a client is Windows 9x (using sendfile from Linux will cause these clients to fail)"

OpenWrt LUCI WebGUI for samba 4 offers 2 tweaks with mentioned properties:

1. ENABLE_EXTRA_TUNING extra tuning options by community feedback
2. DISABLE_ASYNC_IO

[mmakassikis]

use sendfile = yes

the sendfile() syscall improves performance by avoiding copying data back and forth between a the kernel and a process (kernel reads file content from disk, copies it to process buffer, and the process wants to send this buffer to a socket so it is copied back to the kernel)

sendfile() shines when the data should be send as-is, with no further processing or modification.

ksmbd works in the kernel, so there is no copying between kernel-land and userland. However, an additional buffer will be allocated and filled if the client negotiated encryption. If signing is enabled ksmbd has to read the data to compute the signature.

Is authentication configured on ksmbd ?
Are there other configuration differences between ksmbd and samba4 (read/write buffer sizes, max number of credits per connection ...) ?

[bademux]

"ksmbd works in the kernel, so there is no copying between kernel-land and userland"
yep, it is definitely something goes wrong on configuration side

"Is authentication configured on ksmbd ?"
Tested on guest access on fresh OpenWrt 24.10.0.
The same usb disk attached. Cable connected, the same client.

I played with some cache configurations as per https://openwrt.org/docs/guide-user/services/nas/ksmbd#example
configuration below seems to be optimal, but still up to 27MB/s:

	smb2 max read = 1MB
	smb2 max write = 1MB
	smb2 max trans = 1MB

	cache read buffers = no
	cache write buffers = no

[bademux]

Encryption and singing is disabled. Device was restarted after config changed.
Nothing changed. Please find configuration below.

/etc/ksmbd/ksmbd.conf

[global]
	netbios name = Router
	server string = server
	workgroup = WORKGROUP
	interfaces = br-lan 
	bind interfaces only = yes
	ipc timeout = 20
	deadtime = 15
	map to guest = Bad User
	smb2 max read = 1MB
	smb2 max write = 1MB
	smb2 max trans = 1MB
	cache read buffers = no
	cache trans buffers = no
        server signing = disabled
        smb3 encryption = disabled

######### Dynamic written config options #########

[pawstor]
	path = /mnt/storage
	valid users = storage
	create mask = 0666
	directory mask = 0777
	read only = no
	guest ok = yes
	hide dot files = no

[namjaejeon]

Stranged, When you decreased max read/write size, performance is better.

1. How large a file size are you reading?
2. let me know RAM size.
3. Let me know both usb read bandwidth and network bandwidth.

[bademux]

1. Testing on 1GB iso file. It is most common use case for my setup.
2. Router has 256MB ram
   3.1 Connected with 1Gb/s Ethernet cable. Iperf3 shows something >100MB/s
   3.2 USB3 (not uas) connected nvme drive. hdparm -Tt shows ~110MB/s

The same samba4 setup gives stable 38MB/s speeds.

[namjaejeon]

Okay. Thanks.

Can you compare performance of both servers when testing reading only 100MB ?

and check page caches difference after reading file.

[mmakassikis]

can you also try adjusting the max credits parameter ? in the global section:

smb2 max credits = 256

This prevents from queuing too many requests at once (which means more memory allocations, and possibly more fragmentation which ends up reducing performance.)

As Namjae suggested, you should drop caches in between tests to make sure the file is served from storage rather than from RAM:

echo 3 > /proc/sys/vm/drop_caches

(must be run with root privileges)

[bademux]

I applied smb2 max credits = 256 and per dropped ~1MB/s
All tests are done after device reboot, only ksmbd or samba4 installed at the same time to eliminate unwanted possible collisions.
I will also try tests reading only 100MB (As Namjae suggested) and drop cache echo 3 > /proc/sys/vm/drop_caches

I'm not sure about "check page caches difference after reading file". How I can check page cache state on linux system?

[namjaejeon]

I will also try tests reading only 100MB (As Namjae suggested) and drop cache echo 3 > /proc/sys/vm/drop_caches

Your RAM size is small, So I thought there is reclaim overhead to invalidate page caches that read data from disk. So I suggested to read only 100MB.

I'm not sure about "check page caches difference after reading file". How I can check page cache state on linux system?

"use sendfile" seems not to create page cache, i.e. I thought there is no copy overhead. So want to check if there are no page cache in memory. can you give me info (cat /proc/meminfo) about samba and ksmbd after reading file ?

[namjaejeon]

@bademux ksmbd and samba performance is same(or similar) before use_sendfile parameter is enabled ?

[bademux]

@bademux ksmbd and samba performance is same(or similar) before use_sendfile parameter is enabled ?

Yes, it is comparable, about 25-26MB/s.

Sorry for delays, I will try to test page cache stuff asap.

[bademux]

test env information

• before each run caches were dropped with echo 3 > /proc/sys/vm/drop_caches
• the only tested samba impl on installed on the device
• device rebooted after software instalation
• copy done with Nautilus (Gnome file manager)

ksmbd

go ~25MB/s

Router (cat /proc/meminfo)

MemTotal:         249100 kB
MemFree:           91608 kB
MemAvailable:     156508 kB
Buffers:               0 kB
Cached:           109184 kB
SwapCached:            0 kB
Active:            10296 kB
Inactive:         102796 kB
Active(anon):       4244 kB
Inactive(anon):        0 kB
Active(file):       6052 kB
Inactive(file):   102796 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:             0 kB
SwapFree:              0 kB
Dirty:                 0 kB
Writeback:             0 kB
AnonPages:          3968 kB
Mapped:             6196 kB
Shmem:               292 kB
KReclaimable:       2468 kB
Slab:              25420 kB
SReclaimable:       2468 kB
SUnreclaim:        22952 kB
KernelStack:        1016 kB
PageTables:          480 kB
SecPageTables:         0 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:      124548 kB
Committed_AS:      13432 kB
VmallocTotal:    1048372 kB
VmallocUsed:        3796 kB
VmallocChunk:          0 kB
Percpu:              288 kB

Samba4

got ~39MB/s

Router (cat /proc/meminfo)

MemTotal: 249100 kB MemFree: 40240 kB MemAvailable: 147596 kB Buffers: 0 kB Cached: 154460 kB SwapCached: 0 kB Active: 24404 kB Inactive: 137540 kB Active(anon): 11252 kB Inactive(anon): 0 kB Active(file): 13152 kB Inactive(file): 137540 kB Unevictable: 0 kB Mlocked: 0 kB SwapTotal: 0 kB SwapFree: 0 kB Dirty: 0 kB Writeback: 0 kB AnonPages: 7492 kB Mapped: 14512 kB Shmem: 3768 kB KReclaimable: 3696 kB Slab: 26948 kB SReclaimable: 3696 kB SUnreclaim: 23252 kB KernelStack: 1000 kB PageTables: 748 kB SecPageTables: 0 kB NFS_Unstable: 0 kB Bounce: 0 kB WritebackTmp: 0 kB CommitLimit: 124548 kB Committed_AS: 23904 kB VmallocTotal: 1048372 kB VmallocUsed: 3732 kB VmallocChunk: 0 kB Percpu: 288 kB

[bademux]

another thing I can't explain is that testing with console smbcient on ubuntu with 100MB file gives me identical result for smb4 and ksmbd:

smbget -a -W WORKGROUP smb://llrouter.lan/storage/testfile && rm testfile
Using guest user
smb://llrouter.lan/storage/testfile                                             
Downloaded 100,00MB in 4 seconds