>Samba performance issues

Samba performance issues

Comparisons

The Samba server uses TCP to talk to the client. Thus if you are trying to see if it performs well you should really compare it to programs that use the same protocol. The most readily available programs for file transfer that use TCP are ftp or another TCP based SMB server.

If you want to test against something like a NT or WfWg server then you will have to disable all but TCP on either the client or server. Otherwise you may well be using a totally different protocol (such as Netbeui) and comparisons may not be valid.

Generally you should find that Samba performs similarly to ftp at raw transfer speed. It should perform quite a bit faster than NFS, although this very much depends on your system.

Several people have done comparisons between Samba and Novell, NFS or WinNT. In some cases Samba performed the best, in others the worst. I suspect the biggest factor is not Samba vs some other system but the hardware and drivers used on the various systems. Given similar hardware Samba should certainly be competitive in speed with other systems.

Oplocks

Overview

Oplocks are the way that SMB clients get permission from a server to locally cache file operations. If a server grants an oplock (opportunistic lock) then the client is free to assume that it is the only one accessing the file and it will agressively cache file data. With some oplock types the client may even cache file open/close operations. This can give enormous performance benefits.

With the release of Samba 1.9.18 we now correctly support opportunistic locks. This is turned on by default, and can be turned off on a share- by-share basis by setting the parameter :

oplocks = False

We recommend that you leave oplocks on however, as current benchmark tests with NetBench seem to give approximately a 30% improvement in speed with them on. This is on average however, and the actual improvement seen can be orders of magnitude greater, depending on what the client redirector is doing.

Previous to Samba 1.9.18 there was a 'fake oplocks' option. This option has been left in the code for backwards compatibility reasons but it's use is now deprecated. A short summary of what the old code did follows.

Level2 Oplocks

With Samba 2.0.5 a new capability - level2 (read only) oplocks is supported (although the option is off by default - see the smb.conf man page for details). Turning on level2 oplocks (on a share-by-share basis) by setting the parameter :

level2 oplocks = true

should speed concurrent access to files that are not commonly written to, such as application serving shares (ie. shares that contain common .EXE files - such as a Microsoft Office share) as it allows clients to read-ahread cache copies of these files.

Old 'fake oplocks' option - deprecated

Samba can also fake oplocks, by granting a oplock whenever a client asks for one. This is controlled using the smb.conf option "fake oplocks". If you set "fake oplocks = yes" then you are telling the client that it may agressively cache the file data for all opens.

Enabling 'fake oplocks' on all read-only shares or shares that you know will only be accessed from one client at a time you will see a big performance improvement on many operations. If you enable this option on shares where multiple clients may be accessing the files read-write at the same time you can get data corruption.

Socket options

There are a number of socket options that can greatly affect the performance of a TCP based server like Samba.

The socket options that Samba uses are settable both on the command line with the -O option, or in the smb.conf file.

The "socket options" section of the smb.conf manual page describes how to set these and gives recommendations.

Getting the socket options right can make a big difference to your performance, but getting them wrong can degrade it by just as much. The correct settings are very dependent on your local network.

The socket option TCP_NODELAY is the one that seems to make the biggest single difference for most networks. Many people report that adding "socket options = TCP_NODELAY" doubles the read performance of a Samba drive. The best explanation I have seen for this is that the Microsoft TCP/IP stack is slow in sending tcp ACKs.

Read size

The option "read size" affects the overlap of disk reads/writes with network reads/writes. If the amount of data being transferred in several of the SMB commands (currently SMBwrite, SMBwriteX and SMBreadbraw) is larger than this value then the server begins writing the data before it has received the whole packet from the network, or in the case of SMBreadbraw, it begins writing to the network before all the data has been read from disk.

This overlapping works best when the speeds of disk and network access are similar, having very little effect when the speed of one is much greater than the other.

The default value is 16384, but very little experimentation has been done yet to determine the optimal value, and it is likely that the best value will vary greatly between systems anyway. A value over 65536 is pointless and will cause you to allocate memory unnecessarily.

Max xmit

At startup the client and server negotiate a "maximum transmit" size, which limits the size of nearly all SMB commands. You can set the maximum size that Samba will negotiate using the "max xmit = " option in smb.conf. Note that this is the maximum size of SMB request that Samba will accept, but not the maximum size that the *client* will accept. The client maximum receive size is sent to Samba by the client and Samba honours this limit.

It defaults to 65536 bytes (the maximum), but it is possible that some clients may perform better with a smaller transmit unit. Trying values of less than 2048 is likely to cause severe problems.

In most cases the default is the best option.

Locking

By default Samba does not implement strict locking on each read/write call (although it did in previous versions). If you enable strict locking (using "strict locking = yes") then you may find that you suffer a severe performance hit on some systems.

The performance hit will probably be greater on NFS mounted filesystems, but could be quite high even on local disks.

Share modes

Some people find that opening files is very slow. This is often because of the "share modes" code needed to fully implement the dos share modes stuff. You can disable this code using "share modes = no". This will gain you a lot in opening and closing files but will mean that (in some cases) the system won't force a second user of a file to open the file read-only if the first has it open read-write. For many applications that do their own locking this doesn't matter, but for some it may. Most Windows applications depend heavily on "share modes" working correctly and it is recommended that the Samba share mode support be left at the default of "on".

The share mode code in Samba has been re-written in the 1.9.17 release following tests with the Ziff-Davis NetBench PC Benchmarking tool. It is now believed that Samba 1.9.17 implements share modes similarly to Windows NT.

NOTE: In the most recent versions of Samba there is an option to use shared memory via mmap() to implement the share modes. This makes things much faster. See the Makefile for how to enable this.

Log level

If you set the log level (also known as "debug level") higher than 2 then you may suffer a large drop in performance. This is because the server flushes the log file after each operation, which can be very expensive.

Wide lines

The "wide links" option is now enabled by default, but if you disable it (for better security) then you may suffer a performance hit in resolving filenames. The performance loss is lessened if you have "getwd cache = yes", which is now the default.

Read raw

The "read raw" operation is designed to be an optimised, low-latency file read operation. A server may choose to not support it, however. and Samba makes support for "read raw" optional, with it being enabled by default.

In some cases clients don't handle "read raw" very well and actually get lower performance using it than they get using the conventional read operations.

So you might like to try "read raw = no" and see what happens on your network. It might lower, raise or not affect your performance. Only testing can really tell.

Write raw

The "write raw" operation is designed to be an optimised, low-latency file write operation. A server may choose to not support it, however. and Samba makes support for "write raw" optional, with it being enabled by default.

Some machines may find "write raw" slower than normal write, in which case you may wish to change this option.

Read prediction

Samba can do read prediction on some of the SMB commands. Read prediction means that Samba reads some extra data on the last file it read while waiting for the next SMB command to arrive. It can then respond more quickly when the next read request arrives.

This is disabled by default. You can enable it by using "read prediction = yes".

Note that read prediction is only used on files that were opened read only.

Read prediction should particularly help for those silly clients (such as "Write" under NT) which do lots of very small reads on a file.

Samba will not read ahead more data than the amount specified in the "read size" option. It always reads ahead on 1k block boundaries.

Memory mapping

Samba supports reading files via memory mapping them. One some machines this can give a large boost to performance, on others it makes not difference at all, and on some it may reduce performance.

To enable you you have to recompile Samba with the -DUSE_MMAP option on the FLAGS line of the Makefile.

Note that memory mapping is only used on files opened read only, and is not used by the "read raw" operation. Thus you may find memory mapping is more effective if you disable "read raw" using "read raw = no".

Slow Clients

One person has reported that setting the protocol to COREPLUS rather than LANMAN2 gave a dramatic speed improvement (from 10k/s to 150k/s).

I suspect that his PC's (386sx16 based) were asking for more data than they could chew. I suspect a similar speed could be had by setting "read raw = no" and "max xmit = 2048", instead of changing the protocol. Lowering the "read size" might also help.

Slow Logins

Slow logins are almost always due to the password checking time. Using the lowest practical "password level" will improve things a lot. You could also enable the "UFC crypt" option in the Makefile.

Client tuning

Often a speed problem can be traced to the client. The client (for example Windows for Workgroups) can often be tuned for better TCP performance.

See your client docs for details. In particular, I have heard rumours that the WfWg options TCPWINDOWSIZE and TCPSEGMENTSIZE can have a large impact on performance.

Also note that some people have found that setting DefaultRcvWindow in the [MSTCP] section of the SYSTEM.INI file under WfWg to 3072 gives a big improvement. I don't know why.

My own experience wth DefaultRcvWindow is that I get much better performance with a large value (16384 or larger). Other people have reported that anything over 3072 slows things down enourmously. One person even reported a speed drop of a factor of 30 when he went from 3072 to 8192. I don't know why.

It probably depends a lot on your hardware, and the type of unix box you have at the other end of the link.

Paul Cochrane has done some testing on client side tuning and come to the following conclusions:

Install the W2setup.exe file from www.microsoft.com. This is an update for the winsock stack and utilities which improve performance.

Configure the win95 TCPIP registry settings to give better perfomance. I use a program called MTUSPEED.exe which I got off the net. There are various other utilities of this type freely available. The setting which give the best performance for me are:

  1. MaxMTU Remove

  2. RWIN Remove

  3. MTUAutoDiscover Disable

  4. MTUBlackHoleDetect Disable

  5. Time To Live Enabled

  6. Time To Live - HOPS 32

  7. NDI Cache Size 0

I tried virtually all of the items mentioned in the document and the only one which made a difference to me was the socket options. It turned out I was better off without any!!!!!

In terms of overall speed of transfer, between various win95 clients and a DX2-66 20MB server with a crappy NE2000 compatible and old IDE drive (Kernel 2.0.30). The transfer rate was reasonable for 10 baseT.

FIXME The figures are: Put Get P166 client 3Com card: 420-440kB/s 500-520kB/s P100 client 3Com card: 390-410kB/s 490-510kB/s DX4-75 client NE2000: 370-380kB/s 330-350kB/s

I based these test on transfer two files a 4.5MB text file and a 15MB textfile. The results arn't bad considering the hardware Samba is running on. It's a crap machine!!!!

The updates mentioned in 1 and 2 brought up the transfer rates from just over 100kB/s in some clients.

A new client is a P333 connected via a 100MB/s card and hub. The transfer rates from this were good: 450-500kB/s on put and 600+kB/s on get.

Looking at standard FTP throughput, Samba is a bit slower (100kB/s upwards). I suppose there is more going on in the samba protocol, but if it could get up to the rate of FTP the perfomance would be quite staggering.

My Results

Some people want to see real numbers in a document like this, so here they are. I have a 486sx33 client running WfWg 3.11 with the 3.11b tcp/ip stack. It has a slow IDE drive and 20Mb of ram. It has a SMC Elite-16 ISA bus ethernet card. The only WfWg tuning I've done is to set DefaultRcvWindow in the [MSTCP] section of system.ini to 16384. My server is a 486dx3-66 running Linux. It also has 20Mb of ram and a SMC Elite-16 card. You can see my server config in the examples/tridge/ subdirectory of the distribution.

I get 490k/s on reading a 8Mb file with copy. I get 441k/s writing the same file to the samba server.

Of course, there's a lot more to benchmarks than 2 raw throughput figures, but it gives you a ballpark figure.

I've also tested Win95 and WinNT, and found WinNT gave me the best speed as a samba client. The fastest client of all (for me) is smbclient running on another linux box. Maybe I'll add those results here someday ...