I‘ve two servers located in two different data center. Both server deals with a lot of concurrent large file transfers. But network performance is very poor for large files and performance degradation take place with a large files. How do I tune TCP under Linux to solve this problem?
By default the Linux network stack is not configured for high speed large file transfer across WAN links. This is done to save memory resources. You can easily tune Linux network stack by increasing network buffers size for high-speed networks that connect server systems to handle more network packets.
The default maximum Linux TCP buffer sizes are way too small. TCP memory is calculated automatically based on system memory; you can find the actual values by typing the following commands:
$ cat /proc/sys/net/ipv4/tcp_mem
The default and maximum amount for the receive socket memory:
$ cat /proc/sys/net/core/rmem_default
$ cat /proc/sys/net/core/rmem_max
The default and maximum amount for the send socket memory:
$ cat /proc/sys/net/core/wmem_default
$ cat /proc/sys/net/core/wmem_max
The maximum amount of option memory buffers:
$ cat /proc/sys/net/core/optmem_max
Tune values
Set the max OS send buffer size (wmem) and receive buffer size (rmem) to 12 MB for queues on all protocols. In other words set the amount of memory that is allocated for each TCP socket when it is opened or created while transferring files:
# echo 'net.core.wmem_max=12582912' >> /etc/sysctl.conf
# echo 'net.core.rmem_max=12582912' >> /etc/sysctl.conf
You also need to set minimum size, initial size, and maximum size in bytes:
# echo 'net.ipv4.tcp_rmem= 10240 87380 12582912' >> /etc/sysctl.conf
# echo 'net.ipv4.tcp_wmem= 10240 87380 12582912' >> /etc/sysctl.conf
Turn on window scaling which can be an option to enlarge the transfer window:
# echo 'net.ipv4.tcp_window_scaling = 1' >> /etc/sysctl.conf
Enable timestamps as defined in RFC1323:
# echo 'net.ipv4.tcp_timestamps = 1' >> /etc/sysctl.conf
Enable select acknowledgments:
# echo 'net.ipv4.tcp_sack = 1' >> /etc/sysctl.conf
By default, TCP saves various connection metrics in the route cache when the connection closes, so that connections established in the near future can use these to set initial conditions. Usually, this increases overall performance, but may sometimes cause performance degradation. If set, TCP will not cache metrics on closing connections.
# echo 'net.ipv4.tcp_no_metrics_save = 1' >> /etc/sysctl.conf
Set maximum number of packets, queued on the INPUT side, when the interface receives packets faster than kernel can process them.
# echo 'net.core.netdev_max_backlog = 5000' >> /etc/sysctl.conf
Now reload the changes:
# sysctl -p
Use tcpdump to view changes for eth0:
# tcpdump -ni eth0
Recommend readings:
- Please refer to kernel documentation in Documentation /networking/ip-sysctl.txt for more information.
- man page sysctl
🐧 42 comments so far... add one ↓
Category | List of Unix and Linux commands |
---|---|
File Management | cat |
Firewall | Alpine Awall • CentOS 8 • OpenSUSE • RHEL 8 • Ubuntu 16.04 • Ubuntu 18.04 • Ubuntu 20.04 |
Network Utilities | dig • host • ip • nmap |
OpenVPN | CentOS 7 • CentOS 8 • Debian 10 • Debian 8/9 • Ubuntu 18.04 • Ubuntu 20.04 |
Package Manager | apk • apt |
Processes Management | bg • chroot • cron • disown • fg • jobs • killall • kill • pidof • pstree • pwdx • time |
Searching | grep • whereis • which |
User Information | groups • id • lastcomm • last • lid/libuser-lid • logname • members • users • whoami • who • w |
WireGuard VPN | Alpine • CentOS 8 • Debian 10 • Firewall • Ubuntu 20.04 |
great… you always impress me by your level of knowledge and great explanation skills.
it appears that these tuning is not suitable for average workstation users whom want to have a faster/solid network connection?
For average user (Desktop / workstation) defaults are fine. These are for servers with lots of traffic generated while transferring large files.
@Shoaibi: no problem. I’m glade that information helped you to understand new things.
Great article man. Quality stuff you got here… keep em coming.
it will be great if there is another article to light boost network connectivity of workstation/notebook ubuntu.
Hi
According to this page: http://www.psc.edu/networking/projects/tcptune/#Linux
“Kernels greater than 2.6.17 have full auto tuning and manual tuning is generally not recommended.”
If you are in a position to test the effects of auto vs manual tuning thoroughly, please do so.
Thanks
Manual tuning is for experts like this author and myself……. Auto tuning had a very conservative thresholds that are much lower the the a manual tuning can acheve.
Can you offer some specific information about this information:
1) What kernel were you using?
2) Do you have any numbers to show the changes in performance?
3) Are there any negative effects on your systems by doing this?
4) What are the hardware configs of the systems involved?
TNX
OS RHEL 4.7 server – RHEL kernel 2.6.9.xx – 64bit system.
Hardware – Server Dual Core Xeon 53xx with 16GB RAM – RAID 10 with x 15kx 73GB server.
A clear performance decline is observed when using a low value of rmem_max and wmem_max limit used. This causes small window sizes and creates a performance ceiling for large data transfers. However, I lost my data file with trasfer per rate per sec by received size. However, you can easily test things your self try comparison with socket buffer 4 KB and 132 bytes.
I’ve same experience with RHEL 5 64 bit kernel 2.6.18 with 32GB RAM and SAN storage. You need to tweak the kernel.
HTH.
good information, can have any trick to increase wine application’s performance.. i am using tally on wine, it works slow..
Very nice article. I tweaked these parameters a little to obtain almost twice the real throughput in Samba over a high-latency VPN link.
I think this command is a typo:
…should not contain sysctl?
@Sam,
Thanks for the heads-up!
Great info. thanks.
How can I check how much Linux TCP buffer sizes are used on my system?
Interesting, but only close to what I’m looking for. I have a very large number of small connections (client connects via TCP, sends a message that has a maximum size of 64K, and disconnects). The client may reconnect to send an additional message, but EVERY message has it’s own connection. I need to handle >20K (and perhaps over 40K) connections per second over GigE; my server seems to be similar in some ways (32GB of RAM on a SAN, running vanilla Linux 5.2 x86-64, with 2 Quad Core Opterons).
Raising the various memory for individual buffers settings seems futile, since I’m sending large amounts of very small data and would be unlikely to see any performance for mostly unused memory; however,
# echo ‘net.core.netdev_max_backlog = 5000’ >> /etc/sysctl.conf
seems promising, although I think 5K may be low in my case. Any other tuning that might be helpful that you can recommend?
You might want to try a couple of things.. Check the ring buffer sized on your driver using ethtool -g ..
You might want to make sure that your socket is opened using NODELAY flag to turn of nagel and don’t do timestamps=1 and sack=1 .. these things should be turned off to increase efficiency. These things add overhead and usually you don’t need them.
max_backlog depending on the line speed could be set higher than 5000. I used 30000 on a 10gig link but I think this may also be related to the max ring buffer size. Would be interested to know whick buffer the backlog is dumped into.
Some other settings you might want to look at:
net.ipv4.tcp_slow_start_after_idle = 0
and check if there is any interrupt coalescing going on using ethtool -c if your drivre supports it. e1000 supports this. Turn that sucker off Rx/Tx Coalescing..
if you’re sending tons of small messages all the nagel stuff and coalescing will hurt you.
Quite interesting. However, many of the settings are already default on the new Linux kernel (using 2.6.32 on Ubuntu 10.04). I still can’t crack why my Linux workstation with GigE have a transfer rate of 35MBps while my new iMac (OSX) have 65MBps downloading same file from the same NAS using the same port & cable. Is there something I miss?
Are you using CIFS, NFS, or anything else? What about jumbo frame and the
/proc/sys/net/ipv4/tcp_moderate_rcvbuf ? By default most modern Linux kernel comes with autotuning. Also, it depends upon your network card and driver.
I think wmem and rmem are generic for sockets and can be used by any protocol. For tcp it is tcp_wmem and tcp_rmem.
Very Very nice aritlce …Very much helpfull….
Hi Vivek,
I have a small qeury.Please clear me.
As per your word:
“Tune values
Set the max OS send buffer size (wmem) and receive buffer size (rmem) to 12 MB for queues on all protocols. In other words set the amount of memory that is allocated for each TCP socket when it is opened or created while transferring files:”
suppose i set 12 MB for wmem and rmem so that time each tcp socket will take 12 MB ?
means suppose we have bussy server which having 1000 concurrent users so in that time each tcp connection taking 12 MB of ram ?
Please clear me that how to calculate it.
And for such tunning , do we need more ram ?
From where system set 12 mb for each tcp socket ? is from ram ?
Thanks,
Rocky
“suppose i set 12 MB for wmem and rmem so that time each tcp socket will take 12 MB ?
means suppose we have bussy server which having 1000 concurrent users so in that time each tcp connection taking 12 MB of ram ?”
same question. anybody?
Won’t this just cause buffer bloat? http://www.bufferbloat.net/
Hi,
Duplex settings for ethernet interface is often forgotten causing duplex mismatch
and packet drops,
ethtool -s eth0 speed 100 duplex full autoneg off
ethtool eth0
Note : The link goes down and comes up
For GE links autoneg shall be enabled .
The tx queue length of eth0 can also be adjusted using
ifconfig eth0 txqueuelen 3000
Transmit queue length for eth0 can be adjusted
yeah i think it a buffer over bloat , we should not increase the buffer size too much , we should consider the bottleneck on our link while doing such stuff
what is TXqueueing actually
is it recommended on a low latency needed server
–> ifconfig eth0 txqueuelen 3000 <–
after modify,how can i check if the server get better speed?
Hi there, im not sure if I’m suffering this problem, so not sure if this a suitable solution for me. I better explain my case:
I have a little office network, where I have a file, printer, download and router server. I mean I use it as storage (NFS). By a web interface users can access to a mldonkey where they look for things they need to download. And all the network goes to the internet through it.
I recently increased the memory fom 1G to 3G, and added a 2TB hard disk for storage. Since this moment on some workstations, specially the old ones have problems accesing the files stored at the server, if a user needs to watch a video, while watching it, the video stops some times and after a while starts again. The acces to folders with huge amount of data makes the nautilus window to turn gray for a while before showing the content, etc….
These kind of things didn’t happen with 1Gb RAM. I have to say that I have read that large disks may affect performance of the machine. But I suspect my problem is related to ram addition and the network.
I hope you can help me.
Thanks a lot.
Hi, i want to share something that i spent a lot of time to solve and the solution is against any advice found on internet on doing it.
Ambient:
Mini network, 3 computers @ 300Mb wifi, where one is a server (proxy and storage).
To Internet (on server) using wifi direct connected adapter.
Effect:
On file transfer between server and computer, when gets about 50% of bandwidth, the connection with internet slows down to zero bytes transfer.
(The internet adapter is not used to file transfer, its another adapter)
The load of the server is bellow 20%
Server: Fedora Linux kernel 3.6.7-4.fc17.x86_64
Machine 1: Windows 7
Machine 2: Windows 8
This issue has about 18 months, i mean, the time that the network was mounted.
As a experimented developer, I’ve tried really a lot of things and i got really annoyed with this problem and I’ve found a lot of problems related to this found on internet (while searching for a solution). (I was thinking in go to try changes in kernel source, but i was not psychologically prepared to this )
Solution:
The solution was simply increase the network parameters to a very hight values (in memory parameters).
This is NOT proposed to do in a server without a extensive test but, worked like a charm to me. Every network communication had a surprised increase of performance and no more problems.
So, the working parameters are:
# adjust memory to 1.677MB – endless memory :-)
net.core.rmem_max=1677721600
net.core.rmem_default=167772160
net.core.wmem_max=1677721600
net.core.wmem_default=167772160
net.core.optmem_max= 2048000
# set minimum size, initial size, and maximum size in bytes
net.ipv4.tcp_rmem= 1024000 8738000 1677721600
net.ipv4.tcp_wmem= 1024000 8738000 1677721600
net.ipv4.tcp_mem= 1024000 8738000 1677721600
net.ipv4.udp_mem= 1024000 8738000 1677721600
The server is now allocating about 500MB to network buffer and everything is running very well.
Regards
vlassius
Servers to like RAM for caching whatever rubbish you access on them.
Be sure to have a UPS and do proper shutdowns every or else you’re looking at serious corruption and data loss.
Been there, done it, now I have a UPS.
There’s a typo there.
Servers to like RAM -> Servers do like RAM
Works Great, THANKS!
Hi,
Is there any way to reduce the network latency? Want to verify a specific issue, when there is low latency we are get into issue. So to complete testing phase we need to simulate that scenario.
Please advice
Thanks for the article. Should I tune all tcp buffer settings on both sides (i.e. client and server)?
Big thanks. It fixes my server performance issues, such as, TCP Dup Ack and TCP out-of-order problems.
That was great and simple article… The exact way to calculate TCP Buff size:
TCP Buffer Size= *
eg.
20 (ms) * 100 (Mbps) = 0.02 * 100 / 8 * 1024 = 256 KB
Does this required restart of the linux?
How to do the same process for UDP?
I am using UDP for trasfering RTSP stream from server to client over a WiFi. I am also getting a stream on my Client laptop but then after 1 minute the whole network is getting collapsed and it says “No buffer space available” on Server PC.
As there is no network, RTSP stream can not be available.
Please suggest me some solution.
There is an error in net.ipv4.tcp_mem, the value is measured in pages sizes (4k) not in single bytes. I haven’t checked for the udp counterpart.
Useful looking article. Our situation is that we have 10Gb links to two Linux servers, we want them to talk to each other across a 10Gb network; it’s all within the same AS. Our users want to send large amounts of data across the network periodically (eg every day). The tool they are currently using is rsync, and test results are pretty poor. There seem to be two major factors affecting this (taking contention across the shared network for granted) – OpenSSH’s own windowing (for which we are looking at HPN-SSH as a solution), and spikes in traffic which cause TCP sawtoothing . Would you think your solution here would be suitable for our uses?
We are looking to address the contention issues on the backbone with QoS rules.
Use GridFTP
Hello and thanks for a great article.
I wonder if you have any experience with the tc qdisc netem functionality?
I am trying to set very specific delays in the microsecond range, and I need the spread around the set time to be quite accurate. But as it is at the moment we have a bit higher percentage within the delay we have set, and then a long falloff curve with longer delays. So it seems as if there is a buffer that doesn’t empty properly. Is there a way to actually monitor the buffers, send out usage statistics to a file etc..
We are currently using an Emulex oce14102 and the delay is set on a bridge between the two ports of the card. Link speed is 10gbps. If we don’t use netem the bridge itself has good statistics, and when we are sending in just one direction it is also better but you start noticing the falloff curve.