1!== 2!== Speed.txt for Samba release 2.0.10 23 Jun 2001 3!== 4 5Subject: Samba performance issues 6============================================================================ 7 8This file tries to outline the ways to improve the speed of a Samba server. 9 10COMPARISONS 11----------- 12 13The Samba server uses TCP to talk to the client. Thus if you are 14trying to see if it performs well you should really compare it to 15programs that use the same protocol. The most readily available 16programs for file transfer that use TCP are ftp or another TCP based 17SMB server. 18 19If you want to test against something like a NT or WfWg server then 20you will have to disable all but TCP on either the client or 21server. Otherwise you may well be using a totally different protocol 22(such as Netbeui) and comparisons may not be valid. 23 24Generally you should find that Samba performs similarly to ftp at raw 25transfer speed. It should perform quite a bit faster than NFS, 26although this very much depends on your system. 27 28Several people have done comparisons between Samba and Novell, NFS or 29WinNT. In some cases Samba performed the best, in others the worst. I 30suspect the biggest factor is not Samba vs some other system but the 31hardware and drivers used on the various systems. Given similar 32hardware Samba should certainly be competitive in speed with other 33systems. 34 35 36OPLOCKS 37------- 38 39Oplocks are the way that SMB clients get permission from a server to 40locally cache file operations. If a server grants an oplock 41(opportunistic lock) then the client is free to assume that it is the 42only one accessing the file and it will agressively cache file 43data. With some oplock types the client may even cache file open/close 44operations. This can give enormous performance benefits. 45 46With the release of Samba 1.9.18 we now correctly support opportunistic 47locks. This is turned on by default, and can be turned off on a share- 48by-share basis by setting the parameter : 49 50oplocks = False 51 52We recommend that you leave oplocks on however, as current benchmark 53tests with NetBench seem to give approximately a 30% improvement in 54speed with them on. This is on average however, and the actual 55improvement seen can be orders of magnitude greater, depending on 56what the client redirector is doing. 57 58Previous to Samba 1.9.18 there was a 'fake oplocks' option. This 59option has been left in the code for backwards compatibility reasons 60but it's use is now deprecated. A short summary of what the old 61code did follows. 62 63LEVEL2 OPLOCKS 64-------------- 65 66With Samba 2.0.5 a new capability - level2 (read only) oplocks is 67supported (although the option is off by default - see the smb.conf 68man page for details). Turning on level2 oplocks (on a share-by-share basis) 69by setting the parameter : 70 71level2 oplocks = true 72 73should speed concurrent access to files that are not commonly written 74to, such as application serving shares (ie. shares that contain common 75.EXE files - such as a Microsoft Office share) as it allows clients to 76read-ahread cache copies of these files. 77 78Old 'fake oplocks' option - deprecated. 79--------------------------------------- 80 81Samba can also fake oplocks, by granting a oplock whenever a client 82asks for one. This is controlled using the smb.conf option "fake 83oplocks". If you set "fake oplocks = yes" then you are telling the 84client that it may agressively cache the file data for all opens. 85 86Enabling 'fake oplocks' on all read-only shares or shares that you know 87will only be accessed from one client at a time you will see a big 88performance improvement on many operations. If you enable this option 89on shares where multiple clients may be accessing the files read-write 90at the same time you can get data corruption. 91 92SOCKET OPTIONS 93-------------- 94 95There are a number of socket options that can greatly affect the 96performance of a TCP based server like Samba. 97 98The socket options that Samba uses are settable both on the command 99line with the -O option, or in the smb.conf file. 100 101The "socket options" section of the smb.conf manual page describes how 102to set these and gives recommendations. 103 104Getting the socket options right can make a big difference to your 105performance, but getting them wrong can degrade it by just as 106much. The correct settings are very dependent on your local network. 107 108The socket option TCP_NODELAY is the one that seems to make the 109biggest single difference for most networks. Many people report that 110adding "socket options = TCP_NODELAY" doubles the read performance of 111a Samba drive. The best explanation I have seen for this is that the 112Microsoft TCP/IP stack is slow in sending tcp ACKs. 113 114 115READ SIZE 116--------- 117 118The option "read size" affects the overlap of disk reads/writes with 119network reads/writes. If the amount of data being transferred in 120several of the SMB commands (currently SMBwrite, SMBwriteX and 121SMBreadbraw) is larger than this value then the server begins writing 122the data before it has received the whole packet from the network, or 123in the case of SMBreadbraw, it begins writing to the network before 124all the data has been read from disk. 125 126This overlapping works best when the speeds of disk and network access 127are similar, having very little effect when the speed of one is much 128greater than the other. 129 130The default value is 16384, but very little experimentation has been 131done yet to determine the optimal value, and it is likely that the best 132value will vary greatly between systems anyway. A value over 65536 is 133pointless and will cause you to allocate memory unnecessarily. 134 135 136MAX XMIT 137-------- 138 139At startup the client and server negotiate a "maximum transmit" size, 140which limits the size of nearly all SMB commands. You can set the 141maximum size that Samba will negotiate using the "max xmit = " option 142in smb.conf. Note that this is the maximum size of SMB request that 143Samba will accept, but not the maximum size that the *client* will accept. 144The client maximum receive size is sent to Samba by the client and Samba 145honours this limit. 146 147It defaults to 65536 bytes (the maximum), but it is possible that some 148clients may perform better with a smaller transmit unit. Trying values 149of less than 2048 is likely to cause severe problems. 150 151In most cases the default is the best option. 152 153 154LOCKING 155------- 156 157By default Samba does not implement strict locking on each read/write 158call (although it did in previous versions). If you enable strict 159locking (using "strict locking = yes") then you may find that you 160suffer a severe performance hit on some systems. 161 162The performance hit will probably be greater on NFS mounted 163filesystems, but could be quite high even on local disks. 164 165 166SHARE MODES 167----------- 168 169Some people find that opening files is very slow. This is often 170because of the "share modes" code needed to fully implement the dos 171share modes stuff. You can disable this code using "share modes = 172no". This will gain you a lot in opening and closing files but will 173mean that (in some cases) the system won't force a second user of a 174file to open the file read-only if the first has it open 175read-write. For many applications that do their own locking this 176doesn't matter, but for some it may. Most Windows applications 177depend heavily on "share modes" working correctly and it is 178recommended that the Samba share mode support be left at the 179default of "on". 180 181The share mode code in Samba has been re-written in the 1.9.17 182release following tests with the Ziff-Davis NetBench PC Benchmarking 183tool. It is now believed that Samba 1.9.17 implements share modes 184similarly to Windows NT. 185 186NOTE: In the most recent versions of Samba there is an option to use 187shared memory via mmap() to implement the share modes. This makes 188things much faster. See the Makefile for how to enable this. 189 190 191LOG LEVEL 192--------- 193 194If you set the log level (also known as "debug level") higher than 2 195then you may suffer a large drop in performance. This is because the 196server flushes the log file after each operation, which can be very 197expensive. 198 199 200WIDE LINKS 201---------- 202 203The "wide links" option is now enabled by default, but if you disable 204it (for better security) then you may suffer a performance hit in 205resolving filenames. The performance loss is lessened if you have 206"getwd cache = yes", which is now the default. 207 208 209READ RAW 210-------- 211 212The "read raw" operation is designed to be an optimised, low-latency 213file read operation. A server may choose to not support it, 214however. and Samba makes support for "read raw" optional, with it 215being enabled by default. 216 217In some cases clients don't handle "read raw" very well and actually 218get lower performance using it than they get using the conventional 219read operations. 220 221So you might like to try "read raw = no" and see what happens on your 222network. It might lower, raise or not affect your performance. Only 223testing can really tell. 224 225 226WRITE RAW 227--------- 228 229The "write raw" operation is designed to be an optimised, low-latency 230file write operation. A server may choose to not support it, 231however. and Samba makes support for "write raw" optional, with it 232being enabled by default. 233 234Some machines may find "write raw" slower than normal write, in which 235case you may wish to change this option. 236 237READ PREDICTION 238--------------- 239 240Samba can do read prediction on some of the SMB commands. Read 241prediction means that Samba reads some extra data on the last file it 242read while waiting for the next SMB command to arrive. It can then 243respond more quickly when the next read request arrives. 244 245This is disabled by default. You can enable it by using "read 246prediction = yes". 247 248Note that read prediction is only used on files that were opened read 249only. 250 251Read prediction should particularly help for those silly clients (such 252as "Write" under NT) which do lots of very small reads on a file. 253 254Samba will not read ahead more data than the amount specified in the 255"read size" option. It always reads ahead on 1k block boundaries. 256 257 258MEMORY MAPPING 259-------------- 260 261Samba supports reading files via memory mapping them. One some 262machines this can give a large boost to performance, on others it 263makes not difference at all, and on some it may reduce performance. 264 265To enable you you have to recompile Samba with the -DUSE_MMAP option 266on the FLAGS line of the Makefile. 267 268Note that memory mapping is only used on files opened read only, and 269is not used by the "read raw" operation. Thus you may find memory 270mapping is more effective if you disable "read raw" using "read raw = 271no". 272 273 274SLOW CLIENTS 275------------ 276 277One person has reported that setting the protocol to COREPLUS rather 278than LANMAN2 gave a dramatic speed improvement (from 10k/s to 150k/s). 279 280I suspect that his PC's (386sx16 based) were asking for more data than 281they could chew. I suspect a similar speed could be had by setting 282"read raw = no" and "max xmit = 2048", instead of changing the 283protocol. Lowering the "read size" might also help. 284 285 286SLOW LOGINS 287----------- 288 289Slow logins are almost always due to the password checking time. Using 290the lowest practical "password level" will improve things a lot. You 291could also enable the "UFC crypt" option in the Makefile. 292 293CLIENT TUNING 294------------- 295 296Often a speed problem can be traced to the client. The client (for 297example Windows for Workgroups) can often be tuned for better TCP 298performance. 299 300See your client docs for details. In particular, I have heard rumours 301that the WfWg options TCPWINDOWSIZE and TCPSEGMENTSIZE can have a 302large impact on performance. 303 304Also note that some people have found that setting DefaultRcvWindow in 305the [MSTCP] section of the SYSTEM.INI file under WfWg to 3072 gives a 306big improvement. I don't know why. 307 308My own experience wth DefaultRcvWindow is that I get much better 309performance with a large value (16384 or larger). Other people have 310reported that anything over 3072 slows things down enourmously. One 311person even reported a speed drop of a factor of 30 when he went from 3123072 to 8192. I don't know why. 313 314It probably depends a lot on your hardware, and the type of unix box 315you have at the other end of the link. 316 317 318MY RESULTS 319---------- 320 321Some people want to see real numbers in a document like this, so here 322they are. I have a 486sx33 client running WfWg 3.11 with the 3.11b 323tcp/ip stack. It has a slow IDE drive and 20Mb of ram. It has a SMC 324Elite-16 ISA bus ethernet card. The only WfWg tuning I've done is to 325set DefaultRcvWindow in the [MSTCP] section of system.ini to 16384. My 326server is a 486dx3-66 running Linux. It also has 20Mb of ram and a SMC 327Elite-16 card. You can see my server config in the examples/tridge/ 328subdirectory of the distribution. 329 330I get 490k/s on reading a 8Mb file with copy. 331I get 441k/s writing the same file to the samba server. 332 333Of course, there's a lot more to benchmarks than 2 raw throughput 334figures, but it gives you a ballpark figure. 335 336I've also tested Win95 and WinNT, and found WinNT gave me the best 337speed as a samba client. The fastest client of all (for me) is 338smbclient running on another linux box. Maybe I'll add those results 339here someday ... 340 341 342