greg@anacapa.NCEL.Navy.Mil (Gregory K. Ramsey) (10/31/90)
I would like to ask the net on their learned opinions regarding the performance penalty of a 386/486 UNIX vs. DOS. Our situation; we have a division doing AEC type work currently using two 386/33 and several 286 standalone machines running Autocad and some surveying, design and digital terrain modeling packages. We have 300Mbyte 28ms drives and VGA and expect to stay with this for all the new machines. Our idea is to upgrade everyone to 386 or 486 machines and network them with some variety of UNIX & X-windows with one or two servers and 10-15 clients. My main question is what performance degradation (if any) will my users see considering all the overhead that UNIX, X-windows, network communication and system management processes will bring along. Will the 286 (8 & 12 Mhz) users see an apparent performance increase, status quo or decrease. The users perception is what I am really interested in here, but if you have any (or know of any) numbers that would really be appreciated. What about the 386 users? Would the performance be any better if we went with a Novell type net. Any comments on my questions or on any related ideas would be very much appreciated. Post or Email to me and I will summarize to the net if there is interest. Vendors are invited to send me their sales material particularly if it address the questions, but I am particularly interested in the opinions of users. Greg greg@anacapa.ncel.navy.mil Greg Ramsey Naval Civil Engineering Lab Code L54 Port Hueneme, CA 93043 805/982-9720 FAX/982-1418
davidsen@sixhub.UUCP (Wm E. Davidsen Jr) (10/31/90)
In article <652@anacapa.NCEL.Navy.Mil> greg@anacapa.NCEL.Navy.Mil (Gregory K. Ramsey) writes: Forgive the long answer, it's shorter than the question. | I would like to ask the net on their learned opinions regarding | the performance penalty of a 386/486 UNIX vs.DOS. General guidelines, unix vs dos: if you have a job doing a lot of disk i/o it will be somewhat faster on unix due to write back buffering. If your application is compiled for 32 bit unix using a good compiler, and uses 32 bit int data, you will get about 2:1 (I didn't believe it until I measured it either). For programs not using 32 bit data the DOS compilers are slightly better than the unix compilers. Programs which in DOS write directly to the screen are a lot faster than unix screen update, but in many cases, particularly with text modes, the update is faster than you can read the data and it makes no diference. Running DOS programs under UNIX makes screen updates slower. In pixel mode figure VP/ix as 2:1 slower than DOS, DOSmerge 2:1 slower than that. I base this on fractint in 640x480 mode, other program will give other results, but the order of performance is almost always the same. [ they do CAD ] | | We have 300Mbyte 28ms drives and VGA and expect to stay with this | for all the new machines. This is as slow as I would want to go. Spend the money and get 16-18ms drives on the new machines, and you will be happier. | Our idea is to upgrade everyone to 386 or 486 machines and | network them with some variety of UNIX & X-windows with one or | two servers and 10-15 clients. I'm not totally sure what you really mean by that, and I'm not convinced you do either. Remember that the display is the server and the process the client. I presume that you mean you will put 10-15 displays on 1-2 machines. That will give lousy performance compared to DOS. If you mean fileservers, it will save money and system admin time, but hurt performance. You might look at a dedicated NFS server, too. Companies like Epoch have very cost effective storage. | My main question is what performance degradation (if any) will my | users see considering all the overhead that UNIX, X-windows, | network communication and system management processes will bring | along. Given the same number of users (one) performance should be a good bit better. CAD programs really respond to more memory and use of 32 bit data. Running X you should figure 8MB for the system, 2-4MB per user. Stop gasping, memory is down under $50/MB for 1MB SIMM/SIPP. That will help performance more than almost anything else. | Will the 286 (8 & 12 Mhz) users see an apparent performance | increase, status quo or decrease. The users perception is what I | am really interested in here, but if you have any (or know of | any) numbers that would really be appreciated. On a system with adequate memory, you could put two users on a 25MHz 386, and doing the same kinds of things in CAD I would expect they would think they went to heaven. | What about the 386 users? The UNIX version will be faster, but you can load it down. User input is a low load, and you can put a bunch of people on with good response. When they ask for redraw, then you see performance hits for multi-user. Of course I assume you have a 387 in each machine. | Any comments on my questions or on any related ideas would be | very much appreciated. With 486 prices falling I would suggest you go that way. At equal clock speed the 486 is 2-3 times faster than a 386. The cost factor is rarely more than 20% of the total system price (CPU, memory, display, disk, network). You might also look at getting color X_terminals instead of VGA displays. They have higher resolution and will be faster than the builtin display in most if not all cases. -- bill davidsen - davidsen@sixhub.uucp (uunet!crdgw1!sixhub!davidsen) sysop *IX BBS and Public Access UNIX moderator of comp.binaries.ibm.pc and 80386 mailing list "Stupidity, like virtue, is its own reward" -me
peter@ficc.ferranti.com (Peter da Silva) (11/01/90)
In article <2202@sixhub.UUCP> davidsen@sixhub.UUCP (bill davidsen) writes: > If your application is compiled for 32 bit unix using a good > compiler, and uses 32 bit int data, you will get about 2:1 (I didn't > believe it until I measured it either). I'm surprised it's so small. The speed difference for pure-CPU on a 286 between small and large model (accessing 20000-int-long array at effectively random versus accessing a 40000-int-long array) is on the order of 11:1! -- Peter da Silva. `-_-' +1 713 274 5180. 'U` peter@ferranti.com
cpcahil@virtech.uucp (Conor P. Cahill) (11/01/90)
In article <I_S6Q:C@xds13.ferranti.com> peter@ficc.ferranti.com (Peter da Silva) writes: >In article <2202@sixhub.UUCP> davidsen@sixhub.UUCP (bill davidsen) writes: >> If your application is compiled for 32 bit unix using a good >> compiler, and uses 32 bit int data, you will get about 2:1 (I didn't >> believe it until I measured it either). > >I'm surprised it's so small. The speed difference for pure-CPU on >a 286 between small and large model (accessing 20000-int-long array >at effectively random versus accessing a 40000-int-long array) is on >the order of 11:1! Thats because you guys are talking about apples and oranges. The performance difference between small and large models deals with the additional overhead of computing addresses using the segment descriptor (including maintaining the segment descriptors for every far address). The benefits of using a 32 bit compiler on a 32 bit machine deals with using a natural format integer (as far as the processor is concerned) vs having to fake it. -- Conor P. Cahill (703)430-9247 Virtual Technologies, Inc., uunet!virtech!cpcahil 46030 Manekin Plaza, Suite 160 Sterling, VA 22170
@tree.uucp (Chris Gonnerman) (11/02/90)
In article <2202@sixhub.UUCP>, davidsen@sixhub.UUCP (Wm E. Davidsen Jr) writes: [ Bill gives a very clear analysis of DOS vs. Unix on Intel-based boxes ] > -- > bill davidsen - davidsen@sixhub.uucp (uunet!crdgw1!sixhub!davidsen) > sysop *IX BBS and Public Access UNIX > moderator of comp.binaries.ibm.pc and 80386 mailing list > "Stupidity, like virtue, is its own reward" -me The original question had to do with building a network for CAD purposes, and stated that the current setup is 286's running DOS and AutoCAD. If the upgrade machines are 386's, and they buy AutoCAD 386 for DOS, it *supposedly* will run VERY fast. However, it remains to be seen what sort of network you can set up. We are building a similar system (286's, new 386's, and SPARCStation 1's) which will run AutoCAD and NFS/PC-NFS. Our initial research led us to building a system based on DOS on the PC's and SunOS (of course) on the SPARC's. The choice to run DOS was based not on AutoCAD but rather on commercially available Finite Element Analysis software... we were unable to find a really good 386 Unix based software system for this, and a DOS based package specifically for the 386 *is* available. (There may be such software available now, though. We are dealing with government contracting... so the actual specs were written a while back.) In other words, I'm not contradicting Bill, just pointing out that the "best" choice often does not work. The BBS address is because we don't have Internet access where I work... -- +----------------------------------------------------------------------------+ | Chris Gonnerman (Mad Programmer At Large) csusac.ecs.csus.edu!tree!jcg | | @ the Tree BBS, Sacramento, CA ucbvax!ucdavis!csusac!tree!jcg | +---------- DISCLAIMER: These opinions are mine... MINE, I say! -----------+