[Comp.Sci.Dept, Utrecht] Note from archiver<at>cs.uu.nl: This page is part of a big collection of Usenet postings, archived here for your convenience. For matters concerning the content of this page, please contact its author(s); use the source, if all else fails. For matters concerning the archive as a whole, please refer to the archive description or contact the archiver.

Subject: comp.windows.x: Getting more performance out of X. FAQ

This article was archived around: 5 Oct 1998 08:13:25 GMT

All FAQs in Directory: x-faq
All FAQs posted in: comp.windows.x
Source: Usenet Version


Archive-name: x-faq/speedups Last-modified: 1998/09/10 URL: http://www.ualberta.ca/~amulder/speedup-x-faq.txt
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - HOW TO MAXIMIZE THE PERFORMANCE OF X -- monthly posting - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Compiled by Art Mulder (art.mulder@ualberta.ca) More RAM, Faster CPU's, More disk space, Faster Ethernet.... These are the standard responses you hear when you ask how to improve the performance of your workstation. Well, more hardware isn't always an option, and I wonder if more hardware is always even a necessity. This "FAQ" list is a collection of suggestions and ideas from different people on the net on how you can the best possible performance from X Windows on your workstation, WITHOUT PURCHASING MORE HARDWARE. Performance is a highly subjective issue. The individual user must balance `speed' versus `features' in order to come to a personal decision. Therefore this document can be be expected to contain many subjective opinions in and amongst the objective facts. This document is specifically concerned with X. There are of course many other factors that can affect the performance of a workstation. However, they are outside the scope of this document. [ People seriously interested in the whole area of system performance, might want to look at: - "System Performance Tuning" by Mike Loukides. (O'Reilly) - "Configuration and Capacity Planning for Solaris Servers" by Brian Wong (Prentice Hall) - "Sun Performance and Tuning" by A.Cockcroft & R.Petit (Prentice Hall) Or various other books, too numerous to list here. -ed] ----------------- Table of Contents ----------------- 0. Introduction & Administrivia 1. Window Managers 2. The X Server Which Server? Locking the Server into RAM? Starting your Server Fonts About the Resources File Define Your Display Properly 3. Clients A Better Clock for X A Better Terminal Emulator for X Other Client Comments Tuning your client 4. Miscellaneous Suggestions Pretty Pictures A Quicker Mouse Programming Thoughts Backing Store Say What!? * Network Issues (TCP) 5. Other Sources of Information The "Other X FAQ" Books & etc. 6. Author & Notes ! = changed since last issue. * = new since last issue. ----------------------------- Introduction & Administrivia ----------------------------- This document is posted the first monday of each month, to the Usenet news groups comp.windows.x, news.answers, and comp.answers. If you are reading a copy of this FAQ which is more than a few months old (see the "Last-modified" date above) you should probably locate the latest edition, since the information may be outdated. If you do not know how to get those newsgroups and/or your site does not receive them and/or this article has already expired, you can retrieve this FAQ from an archive site. There exist several usenet FAQ archive sites. To find out more about them and how to access them, please see the "Introduction to the *.answers newsgroups" posting in news.answers. The main FAQ archive is at rtfm.mit.edu. This document can be found there in /pub/usenet/news.answers/x-faq/speedups. If you do not have access to anonymous ftp, you can retrieve it by sending a mail message to mail-server@rtfm.mit.edu with the command "send usenet/news.answers/x-faq/speedups" in the message body. In addition, this FAQ is available at ftp.x.org (and its mirror sites) in /contrib/faqs/speedup-x-faq. Most X FAQs are there. --------------- Window Managers --------------- There are a lot of window managers out there, with lots of different features and abilities. The choice of which to use is by necessity a balancing act between performance and useful features. Historically, "twm" was considered to be a good candidate for a speedy window manager. A couple of generic tricks you can try to soup up your window manger, is turning off unnecessary things like "zooming" and "opaque move". Also, if you lay out your windows in a tiled manner, you reduce the amount of cpu power spent in raising and lowering overlapping windows. Joe English (joe@trystero.art.com) I've found that a good font for tiling is 7x13 (aka: -misc-fixed-medium-r-normal--13-100-100-100-c-70-iso8859-1 ). It is the biggest font I know of that I can use on my Sun (1152x900 screen) and still get two 80 column terminal windows side-by-side on the display with no overlap. Other font suggestions will be accepted. Suggestions for your consideration: I used to recommend fvwm here, but the 2.x man page for fvwm now says: "Although it was redesigned to minimize memory consumption, it is now as bloated as any other window-manager. Users searching for a low-memory consumption window-manager are advised to look elsewhere, possibly to an earlier version of fvwm." [thanks to: Ian Thurlbeck (ian@stams.strath.ac.uk) ] You might want to investigate the "Lightweight Window Manager" by Elliott Hughes. http://users.ch.genedata.com/~enh It's reported to be quite spartan. [thanks to: rjarvine@sitriini.hut.fi] ------------ The X Server ------------ Which Server? - - - - - - - Make sure that your server is a proper match for your hardware. If you have a monochrome monitor, use a monochrome X11 server. On my Monochrome Sun, I haven't noticed much difference between the Xsun (colour) server and XsunMono, however it was pointed out to me that XsunMono is about 800k smaller and therefore should contribute to less paging. [ thanks to: Jonny Farringdon (j.farringdon@psychol.ucl.ac.uk), Michael Salmon (Michael.Salmon@eos.ericsson.se) ] How your server was compiled can also make a difference. Jeff Law (law@schirf.cs.utah.edu) advises us that on a Sun system, X should be compiled with gcc (version 2.*) or with the unbundled Sun compiler. You can expect to get "*very* large speedups in the server" by not using the bundled SunOS compiler. I assume that similar results would occur if you used one of the other high-quality commercial compilers on the market. Ben Jackson (bjj@sequent.com): The XFree86 XF86_SVGA server that comes with the binary distributions has support for 17 different kinds of SVGA hardware. On any single machine installation, support for all but one can be removed, drastically reducing the size of the server binary. This is accomplished by installing the LinkKit distribution (available from ftp.xfree86.org, or the mirror site where you got the rest of the tarballs) and editing the included `site.def' definition of XF86SvgaDrivers. More details are included in the README (installs as /usr/X11R6/lib/Server/README). On my FreeBSD 2.0.5 machine, the stock XF86_SVGA is 2.7M. A cirrus-only version is 1.5M, and took less than 5 minutes to build. Locking the Server into RAM? - - - - - - - - - - - - - - - It was suggested that perhaps hacking the X server so that it would stay locked in RAM (and therefore not get paged out) might help performances. eg: via a call to plock(). [Eric C Claeys (ecc@eperm.att.com), Danny Backx (db@sunbim.be), Juan D. Martin (juando@cnm.us.es)] Kenny Ranerup [kenny@axis.se] tried this a few years ago and confirmed that it does work, and can give a noticeable speedup. But, you must be careful that your server doesn't grow too large (ie, close to the size of your RAM). He also makes the point that slow interactive performace is in many cases due to client paging (paging in libraries), and not server paging. Starting your Server - - - - - - - - - - - Joe English (joe@trystero.art.com) : If you start up a lot of clients in your .xsession or whatever, sleep for a second or two after launching each one. After I changed my .xclients script to do this, logging in actually took *less* time... we have a heavily loaded system without much core, though. This sounds crazy, but I have confirmed that it works! Warner Losh (imp@boulder.openware.com) provided me with a good explanation of why this works, which I have summarized here: When you start up an X server it takes a huge amount of time to start accepting connections. A lot of initialization is done by the server when it starts. This process touches a large number of pages. Any other process running at the same time would fight the server for use of the CPU, and more importantly, memory. If you put a sleep in there, you give the Server a chance to get itself sorted out before the clients start up. Similarly, there is also a lot of initialization whenever an X client program starts: toolkits registering widgets, resources being fetched, programs initializing state and "databases" and so forth. All this activity is typically memory intensive. Once this initialization is done ("The process has reached a steady state"), the memory usage typically settles down to using only a few pages. By using sleeps to stagger the launching of your clients in your .Xinitrc , you avoid them fighting each other for your workstation's limited resources This is most definitely a "Your Mileage May Vary" situation, as there are so many variables to be considered: available RAM, local swap space, load average, number of users on your system, which clients you are starting, etc. Currently in my .xinitrc I have a situation like: (sleep 1; exec xclock ) & (sleep 1; exec xbiff ) & (sleep 1; exec xterm ) & (sleep 1; exec xterm ) & I've experimented with: (sleep 1; exec xclock ) & (sleep 2; exec xbiff ) & (sleep 3; exec xterm ) & (sleep 4; exec xterm ) & I've even tried: (sleep 2; exec start_X_clients_script ) & and then in start_X_clients_script I had: (sleep 1; exec xclock ) & (sleep 1; exec xbiff ) & (sleep 1; exec xterm ) & (sleep 1; exec xterm ) & [ The idea with this last one was to make sure that xinit had completely finished processing my .xinitrc, and had settled down into a "steady state" before the sleep expired and all my clients were launched. ] All of these yielded fairly comparable results, and so I just stuck with my current setup, for its simplicity. You will probably have to experiment a bit to find a setup which suits you. Theo Honohan <th208@cam.ac.uk> suggests that you might want to investigate 'xtoolwait' (which can be found at ftp.x.org/contrib/utilities/xtoolwait-1.1.tar.gz) which does similar pacing of X startup. Fonts - - - Loading fonts takes time and RAM. If you minimize the number of fonts your applications use, you'll get speed increases in load-up time. One simple strategy is to choose a small number of fonts (one small, one large, one roman, whatever suits you) and configure all your clients -- or at least all your heavily used clients -- to use only those few fonts. Client programs should start up quicker if their font is already loaded into the server. This will also conserve server resources, since fewer fonts will be loaded by the server. [ Farrell McKay (fbm@ptcburp.ptcbu.oz.au), Joe English (joe@trystero.art.com) ] eg: My main xterm font is 7x13, so I also have twm set up to use 7x13 in all its menus and icons etc. Twm's default font is 8x13. Since I don't normally use 8x13, I've eliminated one font from my server. Oliver Jones (oj@roadrunner.pictel.com): Keep fonts local to the workstation, rather than loading them over nfs. If you will make extensive use of R5 scalable fonts, use a font server. Anthony Stuckey (astuckey@predator.urbana.mcd.mot.com): This is particularly good advice on Linux. Linux has a problem with non-floating-point capable hardware (486SX's, etc) in that they will freeze for up to a few minutes while the Scalable fonts are created... The font server fixes this to a certain extent, as does removing the scalable fonts from your font path. ... I believe that a writeup of this problem is in the Linux Tips FAQ [The main point here is that non-fpu machines, linux or otherwise, should experience a benefit from running a font server.] About the Resources File - - - - - - - - - - - - - Keep your .Xresources / .Xdefaults file small. Saves RAM and saves on server startup time. Joe English (joe@trystero.art.com) One suggestion: In your .Xresources file, try putting only the minimum number of resources that you want to have available to all of your applications. For example: *reverseVideo: true Then, separate your resources into individual client-specific resource files. For example: $HOME/lib/app-defaults. In your .login file set the environment variable XUSERFILESEARCHPATH: setenv XUSERFILESEARCHPATH $HOME/lib/app-defaults/%N (Note: It is easier and quicker for Xt to process XUSERFILESEARCHPATH than XAPPLRESDIR.) [ The "comp.windows.x Frequently Asked Questions" FAQ contains an excellent explanation of how this and the other main X environment variables work. --ed.] So, when xterm launches, it loads (through Xt) its resources from .../app-defaults/XTerm. Xdvi finds them in .../app-defaults/XDvi, and so on and so forth. Note that not all clients follow the same XXxxx resource-file naming pattern. You can check in your system app-defaults directory (often: /usr/X11R5/lib/X11/app-defaults/) to find the proper name, and then name your personal resource files with the same name. This is all documented in the Xt Specification (pg 125 & 666). [Thanks to: Kevin Samborn (samborn@mtkgc.com), Michael Urban (urban@cobra.jpl.nasa.gov), Tom Bagli (tomb@gator.bocaraton.ibm.com), and Mike Long (mikel@ee.cornell.edu). Kevin is willing mail his setup files to inquirers.] This method of organizing your personal resources has the following benefits: - Easier to maintain / more usable. - Fewer resources are stored in the X server in the RESOURCE_MANAGER property. As a side benefit your server may start fractionally quicker, since it doesn`t have to load all your resources. - Applications only process their own resources, never have to sort through all of your resources to find the ones that affect them. It also has drawbacks: - the application that you are interested in has to load an additional file every time it starts up. This doesn't seem to make that much of a performance difference, and you might consider this a huge boon to usability. If you are modifying an application's resource database, you just need to re-run the application without having to "xrdb" again. - xrdb will by default run your .Xresources file through cpp. When your resources are split out into multiple resource files and then loaded by the individual client programs, they will not. WATCH OUT FOR THIS!! I had C style comments in my .Xresources file, which cpp stripped out. When I switched to this method of distributed resource files I spent several frustrating days trying to figure out why my clients were not finding their resources. Xt did *NOT* provide any error message when it encountered the C style comments in the resource files, it simply, silently, aborted processing the resource file. The loss of preprocessing (which can be very handy, e.g. ``#ifdef COLOR'' ...) is enough to cause some people to dismiss this method of resource management. - You may also run into some clients which "break the rules". For example, neither Emacs (18.58.3) nor Xvt (1.0) will find their resources if they are anywhere other than in .Xresources. This is because emacs (and presumably, Xvt) are not Xt based applications. They are coded to use Xlib directly. XUSERFILESEARCHPATH and XAPPLRESDIR are Xt features. Larry Williamson (larry@witch.mitra.com) - when starting up a client on a machine that does not share files with the machine where your resources are stored, your client will not find its resources. Loading all your resources into the server will guarantee that all of your clients will always find their resources. Casey Leedom (casey@gauss.llnl.gov) A possible compromise that I have tried, is to put resources for all my heavily used clients (eg: xterm) into my .Xresources file, and to use the "separate resources files" method for clients that I seldom use. Define Your Display Properly - - - - - - - - - - - - - - - Client programs are often executed on the same machine as the server. In that situation, rather than setting your DISPLAY environment variable to "<hostname>:0.0", where <hostname> is the name of your workstation, you should set your DISPLAY variable to "unix:0.0" or ":0.0". By doing this you access optimized routines that know that the server is on the same machine and use a shared memory method of transferring requests. [thanks to Patrick J Horgan (pjh70@ras.amdahl.com)] See the _DISPLAY NAMES_ section of the X(1) man page for further explanation of how to properly set your display name. "I don't think it's stock MIT, but (at least) Data General and HP have libraries that are smart enough to use local communication even when the DISPLAY isn't set specially." Rob Sartin (88opensi!sartin@uunet.UU.NET) [Jody Goldberg (jody@algorithmics.com) sent me an Xlib patch to change stock R5 to use local communication even if DISPLAY is not properly set. I don't want to get in the business of distributing or trying to juggle non-MIT patches and so have elected not to include it here. Hopefully MIT will apply this minor (~8 lines) patch themselves. In the meantime, if you want to try it yourself, email Jody. --ed.] ------- Clients ------- If you only have a few megabytes of Ram then you should think carefully about the number of programs you are running. Think also about the _kind_ of programs you are running. For example: Is there a smaller clock program than xclock? Unfortunately, I haven't really noticed that programs advertise how large they are, so the onus is on us to do the research and spread the word. [ Suggestions on better alternatives to the some of the standard clients (eg: Xclock, Xterm, Xbiff) are welcome. --ed.] I've received some contradictory advice from people, on the subject of X client programs. Some advocate the use of programs that are strictly Xlib based, since Xt, Xaw and other toolkits are rather large. Others warn us that other applications which you are using may have already loaded up one or more of these shared libraries. In this case, using a non-Xt (for example) client program may actually _increase_ the amount of RAM consumed. The upshot of all this seems to be: Don't mix toolkits. That is, try and use just Athena clients, or just Xview clients (or just Motif clients, etc). If you use more than one, then you're dragging in more than one toolkit library. Know your environment, and think carefully about which client programs would work best together in that environment. [Thanks to: Rob Sartin (88opensi!sartin@uunet.UU.NET), Duncan Sinclair (sinclair@dcs.gla.ac.uk | sinclair@uk.ac.gla.dcs) ] A Better Clock for X - - - - - - - - - - - 1) xcuckoo suggested by: Duncan Sinclair (sinclair@dcs.gla.ac.uk) available: on export.lcs.mit.edu Xcuckoo displays a clock in the title bar of *another* program. Saves screen real estate. 2) Swisswatch suggested by: Theo Honohan <th208@cam.ac.uk ftp://sunsite.doc.ic.ac.uk/packages/X11/R5contrib/swisswatch-0.06.tar.Z Of course, the ultimate clock --- one that consumes no resources, and takes up no screen real estate --- is the one that hangs on your wall. :-) A Better Terminal Emulator for X - - - - - - - - - - - - - - - - - From the README file distributed with xterm: +----- | Abandon All Hope, Ye Who Enter Here | | This is undoubtedly the most ugly program in the distribution. | ... +----- Ugly maybe, but at my site it's still the most used. I suspect that xterm is one of the most used clients at many, if not most sites. Laziness? Isn't there a better terminal emulator available? See below. If you must use xterm, you can try reducing the number of saveLines to reduce memory usage. [ Oliver Jones (oj@roadrunner.pictel.com), Jonny Farringdon (j.farringdon@psychol.ucl.ac.uk) ] 1) Xvt Suggested by: Richard Hesketh (rlh2@ukc.ac.uk) Author: John D. Bovey (jdb@ukc.ac.uk) Available: http://unix.hensa.ac.uk/pub/misc/unix/xvt [From the README file] "Xvt is an X terminal-emulator that is designed to be more or less compatible with xterm while using much less swap space. It is mainly intended for use at sites which use large numbers of X terminals but may also be useful on single workstations that are short of memory. On a [SunOS 4.1.* Sparc], an initially invoked xvt uses about 1/3 megabyte of swap while xterm uses about 1.3 megabytes (obtained by running pstat rather than ps which seems to give unreliable size figures on SPARCs). The main way that xvt achieves its small size is by avoiding the use of the X toolkit." There have been some conflicting opinions aired over xvt vs. xterm. Different sites with different needs will likely have to do their own evaluation. Caveat Emptor, your mileage may vary. 2) rxvt Available: http://babayaga.math.fu-berlin.de:80/~rxvt/ Originally a stripped down Linux port of xvt, with some minor modifications, it's now being actively developed on it's own. [thanks to: Theo Honohan <th208@cam.ac.uk for the update] Other Client Comments - - - - - - - - - - - When using the xlock screen locking client, the "qix" mode is much more lightweight than the default mode. [Q. Alex Zhao (azhao@cc.gatech.edu)] I can confirm the previous report from personal experience here. In a lab full of ~20 sun 3/80's all running the Xkernel software, the subnet was getting regularly clobbered. This was tracked down to students running xlock. Since these machines are all xterminals, the actual xlock program runs remotely, and generates a scary amount of network traffic to draw the default 'swarm' pictures on the screen. We have removed this program and replaced it with a much simpler screen locking program which simply blanks the screen. [ed.] A suggestion: xtrlock, by Ian Jackson (ijackson@nyx.cs.du.edu) It is very small (about 16k running), and doesn't produce any traffic while locked. Drawback: It doesn't make your screen invisible. [ Klamer Schutte (klamer@ph.tn.tudelft.nl) ] An improved xbiff: Xbuffy. It can watch multiple mailboxes and can do a pop up showing the From and Subject line of incoming mail. Xbiff doesn't eat much memory and Xbuffy tends to be about the same (on the RS/6000, xbuffy tends use just a little less memory). It is in the contrib directory on ftp.x.org. [ Bill Pemberton (wfp5p@virginia.edu) ] Tuning your client - - - - - - - - - - Suggestions on how you can tune your client programs to work faster. From Scott Barman (scott@asd.com) comes a suggestion regarding Motif Text Field Widgets: I noticed that during data entry into Motif text field widgets, I was getting a slight lag in response to some keystrokes, particularly the initial one in the field. Examining the what was going on with xscope I found it. It seems that when the resource XmNblinkRate is non-zero and the focus is on a text field widget (or even just a text widget) the I-beam cursor will blink. Every time the cursor appears or disappears in those widgets, the widget code is making a request to the server (CopyArea). The user can stop this by setting the resource XmNblinkRate to 0. It is not noticeable on a 40MHz SPARC, but it does make a little difference on a [slower system]. This specific suggestion can probably be applied in general to lots of areas. Consider your heavily used clients, are there any minor embellishments that can be turned off and thereby save on Server requests? ------------------------- Miscellaneous Suggestions ------------------------- Pretty Pictures - - - - - - - - Don't use large bitmaps (GIF's, etc) as root window backgrounds. - The more complicated your root window bitmap, the slower the server is at redrawing your screen when you reposition windows (or redraw, etc) - These take up RAM, and CPU power. I work on a Sun SPARC and I'm conscious of performance issues, I can't comprehend it when I see people with a 4mb Sun 3/60 running xphoon as their root window. I'll let someone else figure out how much RAM would be occupied by having a full screen root image on a colour workstation. - If you're anything like me, you need all the screen real estate that you can get for clients, and so rarely see the root window anyway. [ Thanks to Qiang Alex Zhao (azhao@cs.arizona.edu) for reminding me of this one. --ed.] A Quicker Mouse - - - - - - - - Using xset, you can adjust how fast your pointer moves on the screen when you move your mouse. I use "xset m 3 10" in my .xinitrc file, which lets me send my pointer across the screen with just a flick of the wrist. See the xset man page for further ideas and information. Hint: sometimes you may want to *slow down* your mouse tracking for fine work. To cover my options, I have placed a number of different mouse setting commands into a menu in my window manager. e.g. (for twm) : menu "mouse settings" { "Mouse Settings:" f.title " Very Fast" ! "xset m 7 10 &" " Normal (Fast)" ! "xset m 3 10 &" " System Default (Un-Accelerated)" ! "xset m default &" " Glacial" ! "xset m 0 10 &" } Programming Thoughts - - - - - - - - - - - Joe English (joe@trystero.art.com) : To speed up applications that you're developing, there are tons of things you can do. Some that stick out: - For Motif programs, don't set XmFontList resources for individual buttons, labels, lists, et. al.; use the defaultFontList or labelFontList or whatever resource of the highest-level manager widget. Again, stick to as few fonts as possible. - Better yet, don't use Motif at all. It's an absolute pig. - Don't create and destroy widgets on the fly. Try to reuse them. (This will avoid many problems with buggy toolkits, too.) - Use a line width of 0 in GCs. On some servers this makes a HUGE difference. - Compress and collapse multiple Expose events. This can make the difference between a fast application and a completely unusable one. Francois Staes (frans@kiwi.uia.ac.be) : Just a small remark: I once heard that using a better malloc function would greatly increase performance of Xt based applications since they use malloc heavily. They suggested trying out the GNUY malloc, but I didn't find the time yet. I did some tests on small programs just doing malloc and free, and the differences were indeed very noticeable ( somewhat 5 times faster) [ Any confirmation on this from anyone? --ed.] Andre' Beck (Andre_Beck@IRS.Inf.TU-Dresden.de) : - Unnecessary NoExpose Events. Most people use XCopyArea/XCopyPlane as fastest blit routines, but they forget to reset graphics_exposures in the GC used for the blits. This will cause a NoExpose Event every blit, that, in most cases, only puts load onto the connection and forces the client to run through its event-loop again and again. - Thousands of XChangeGC requests. This "Gfx Context Switching" is also seen in most handcoded X-Apps, where only one or few GCs are created and then heavily changed again and again. Xt uses a definitely better mechanism, by caching and sharing a lot of GCs with all needed parameters. This will remove the load of subsequent XChangeGC requests from the connection (by moving it toward the client startup phase). Backing Store - - - - - - - Joe Nardelli (nardelli@mitre.org) : I've found that using 'backing store' really speeds up redrawing of complex windows. Ex: I have a window that stores an intricate 3-D sceen, and redrawing it every time a window pops up in front is a real waste of time. It is much quicker to set the 'backing store' attribute to either 'Always' or 'WhenMapped'. ... Note that the server keeps a copy of the window in memory, so very simple windows may actually slow some systems done. Your mileage may vary. Here is another way of looking at the issue of backing store. If you are low on memory, you should disable backing store, since a redraw will be quicker than pulling the saved image from swap space. (We are assuming here that you are sufficiently low on memory that your backing store ends up in swap space) If you have enough memory, backing store will beat redraw all the time -- unless you have _very_ fast hardware. [from discussions with Nicolai Langfeldt (janl@ifi.uio.no) ] Say What!? - - - - - - Some contributors proposed ideas that seem right off the wall at first: David B. Lewis (by day: dbl@osf.org, by night: david%craft@uunet.uu.net) : How about this: swap displays with someone else. Run all your programs on the other machine and display locally; the other user runs off your machine onto the other display. Goal: reduce context switches in the same operation between client and server. I'm not in a situation where I can easily try this, but I have received the following confirmation... Michael Salmon (Michael.Salmon@eos.ericsson.se): I regularly run programs on other machines and I notice a big difference. I try to run on a machine where I will reduce net usage and usually with nice to reduce the impact of my intrusion. This helps a lot on my poor little SS1+ with only 16 MB, it was essential when I only had 8 MB. Casey Leedom (casey@gauss.llnl.gov) : [The X11 Server and the client are] competing for the same CPU as your server when you run it on the same machine. Not really a major problem, except that the X11 client and the server are in absolute synchronicity and are context thrashing. Timothy H Panton (thp@westhawk.uucp) : Firstly it relies on the fact that most CPU's are mostly idle, X's cpu usage is bursty. so the chances of you and your teammate doing something cpu-intensive at the same time is small. If they are not then you get twice the cpu+memory available for your action. The second factor is that context switches are expensive, using 2 cpu's halves them, you pay a price due to the overhead of going over the network, but this is offset in most cases by the improved buffering of a network (typically 20k vs 4k for a pipe), allowing even fewer context switches. Network Issues (TCP) - - - - - - - - - - - It is questionable whether the following belongs in this FAQ, since it seems to mostly be a question of network performances. But here it is: Bob Arendt (rdarendt@mcmail.com) : It turns out that matching TCP sendspace and recvspace on client and server can dramatically affect performance. Sun/UNIX and TCPware/VMS default to 4Kbyte buffer sizes (out-of-the-box). Ultrix/DEC and DigitalUnix/DEC default to 32Kbyte buffer sizes. Using the same workstation hardware and FDDI network connections I got the following stats: x11perf -putImage100 (run on TCPware/VMS client, default 4K) = 4.5 msec with TCPware/VMS server, default 4K = 600 msec with Ultrix/VMS server, default 32K = 600 msec with DigitalUnix server, default 32k Changing the Ultrix config and DigitalUnix kernal to 4K TCP send/recv sizes resulted in matching speeds of 4.5 msec. In this case, less was more :-) X11 must use system TCP default buffer settings. With a mismatched, large recvs pace, it appears that polling mechanism (ala select()?) empties the buffers rather than triggering on buffer-full. Matching to small buffers is probably better than larger buffers since it tends to reduce latency and most network devices break up traffic into smaller chunks anyway. This also appears to be an issue for Windows/X11 servers. Running windowsNT on the same Dec Alpha with a third-party X11 server I see the same dismal response times. I suspect it's the same problem (in conjuncution with X11->winDoze translation delay). Don't know how to tune windowsNT though. ---------------------------- Other Sources of Information ---------------------------- The "Other X FAQ" - - - - - - - - - David B. Lewis (faq%craft@uunet.uu.net) maintains the informative and well written "comp.windows.x Frequently Asked Questions" document. Its focus is on general X information, while this FAQ concentrates on performance. The comp.windows.x FAQ does address the issue of speed, but only with regards to the X server. The gist of that topic seems to be: "Use X11R5, it is faster than R4". (Please see the X FAQ for complete details). Books & etc. - - - - - - - Volume 8 in O'Reilly's X Window System Series; ``X Window System Administrator's Guide'' is a book that all X administrator's should read. (email your snailmail address to catalog@ora.com for a catalog) Adrian Nye (adrian@ora.com): A lot more tips on performance are in the paper "Improving X Application Performance" by Chris D. Peterson and Sharon Chang, in Issue 3 of The X Resource. An earlier version of this paper appeared in the Xhibition 1992 conference proceedings. This paper is absolutely essential reading for X programmers. [For information on The X Resource, contact paula@ora.com --ed.] Sajee Mathew (scm@cdssua.chesapeake.com) "Motif Debugging and Performance Tuning" by Doug Young, Prentice Hall, 1994. Good book. The section on performance tuning presents some useful techniques for speeding up X programs from the X-lib and toolkit levels. -------------- Author & Notes -------------- This list is currently maintained by Art Mulder (art.mulder @ ualberta.ca) Suggestions, corrections, or submission for inclusion in this list are gladly accepted. Layout suggestions and comments (spelling mistak's too! :-) are also welcome. Currently I have listed all contributors of the various comments and suggestions. If you do not want to be credited, please tell me. speedup-x-faq is copyright (c) 1993-96 by Arthur E. Mulder You may copy this document in whole or in part as long as you don't try to make money off it, or pretend that you wrote it. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - PLEASE NOTE: The reply address in the headers is bogus. I regret this action, but it is necessary to ATTEMPT to defeat Junk Email robots. -- ...art mulder ( art.mulder@ualberta.ca )( http://www.ualberta.ca/~amulder/ ) ( Sys Admin / Support Analyst, Network Resources ) ( Computing and Network Services, U of Alberta, Edmonton )