Aucbvax.5834 fa.works utcsrgv!utzoo!decvax!ucbvax!works Mon Jan 18 13:47:17 1982 WorkS Digest V2 #7 >From JSOL@USC-ECLB Sat Jan 16 19:51:09 1982 Works Digest Saturday, 16 Jan 1982 Volume 2 : Issue 7 Today's Topics: Virtual Memory Concepts, Memory Schemes ---------------------------------------------------------------------- Date: 14 January 1982 01:16 est From: SSteinberg.SoftArts at MIT-Multics Subject: WORKS V2 #4: Memory Schemes 1. Why virtual memory? When the large computers came out everyone said that memory was getting cheaper and that virtual memory roached performance and was obviously unnecessary. Everyone learned to lived in cramped memory space and thousands of programmer hours were wasted on hairbag overlay schemes and disk based memory managers. When minicomputers came out everyone said that memory was getting cheaper and that virtual memory roached performance and was obviously unnecessary. Everyone went bananas trying to cram anything into itty bitty 64K (B or W) address spaces which was especially ludicrous since the machine often had .5M(B/W) on it. When midicomputers came out I personally flamed at a couple of the manufacturers suggesting that they consider putting in VM. They at least conceded that it could be done efficiently but they couldn't imagine why I'd need given a full 20 bit address space. When I explained that I would have to limit the size of the symbol table at x, they explained that I should write a software paging system. When microcomputers came out everyone was blown away by the fact that you could stack 300 of them in the carcass of a bombardier beetle and have room left over for an I/O multiplexer chip so no one considered VM. So what do we have today? "Why can't VisiCalc hold more stuff in it?" "Damned, the editor buffer is full again and now I can't just add memory." "You can't call a Macsyma program from a PL/I program because they don't both fit in memory even if there is enough room." In other words. Everyone is bumping into the same garbage they bumped into ten years ago except that more people are bumping into it and this time we might all win. As far as segmentation goes I approve completely. Segmentation lets PL/I call Macsyma since it enforces a set of rules which describe how all programs have to communicate and it allows them to share an address space (segmentation is not as important on object oriented machines since each object can be viewed as a segment). A very successful segmented machine is the HP41C calculator. It provides segmentation and dynamic linking in a tangible form. It provides one system segment, one writable segment and four user addable segments each of which may contain any number of named entry points. A key may be bound to any entry point. When I buy a device, such as the magnetic card reader, I plug it into one of the segment slots and suddenly I can invoke a whole series of programs which can read and write data to this device. If I buy the business package I can suddenly run the TBOND function to evaluate Treasury bills and so on. I am often surprised that such a simple and powerful scheme is used so rarely. ------------------------------ Date: 14 Jan 1982 08:19:14-PST From: mo at LBL-UNIX (Mike O'Dell [system]) Subject: Another view of virtual memory While the comments about a person's desires exceeding his machine is a good argument for virtual memory, I would like to present a somewhat different view. (50 pounds of code in a 5 pound machine is the classical party line taken by many vendors who sell machines with virtual memory (except Burroughs).) I view paging as primarily a physical memory management and protection strategy. While real memory will never be large enough for everything you can think of, it is large enough so a bit here and there can be sacrificed to fragmentation. Relocating and protecting access of things via the hardware in nice, fixed-size chunks makes things much easier. Anyone who has written a memory allocator knows it is MUCH easier to write a good one for fixed-sized blocks than one for variable requests. Not having to scrimp and save each byte, and the availability of the paging hardware makes managing the contents of physical memory much easier. It also lets you play the 50 pounds of code game if you want, but that is not as visibly important (I realize it is to MANY people). I have a more Multics-oriented view of segments - they are structuring tools (code or data). However, I do subscribe to the "1-1/2 dimensional" virtual memory schemes. By 1-1/2 dimensions I mean the bits carry from page into segment, rather than the Multics 2 dimensional scheme whereby they don't. There are several reasons. Unless you use gigantic addresses (like the HP scheme mentioned here before), you can never have the right number of bits where you need them. Some programs have huge arrays so need very large segments, others have many arrays, needing many segments, and still others have both!! While the S1 scheme of a floating boundary is clever, it is hard to do if you aren't designing your own hardware. Therefore, for machines with a flat virtual address space, like the VAX (which are vexed by a small page size), a good scheme is 16 bits of page and 16 bits of segment with carry's allowed. This allows a file in the filesystem, or a large array, to be mapped into memory as several contiguous segments, providing consecutive, indexable addresses. To prevent tragedy, simply force the following segment to be a hole which will access trap. If you put these firewalls between each active segment, you do reduce the maximum number of distict, active segments to 2**15, but that is still quite a few (I know, Multics-ers wouldn't be satisfied). And if the segments are code, they can normally be allocated adjacently with reasonable safety. The 1-1/2D scheme is straightforward to implement given a machine with only demand paging, enough logical address bits (32!), and a decent page size. The VAX bairly qualifies on this last point. To really use the VAX 2 gigabyte logical address space requires 8 megabytes of physical page tables, if you don't share segment tables. This is the current maximum physical memory!!!! People building pagers should go read the studies IBM did a long time ago. The picked 1K and 4K byte page sizes for a good reason!! -Mike ------------------------------ Date: 14 January 1982 12:43-EST From: Stavros M. Macrakis Subject: WorkS Digest V2 #5: VM Large address space gives you uniformity of reference, as has been pointed out, across physical configurations and for that matter software configurations (what happens to your 'real memory' when you need to enlarge some process's workspace?). The penalty is presumed to be performance. The discussion on virtual memory so far has assumed that swap management is independent of the application. It would appear an unwarranted assumption. When performance requires it, it is normal to tweak lower-level virtual machines in one's hierarchy. For instance, if it is discovered that some particular kind of array calculation is taking a large fraction of the runtime of an important program, one may well wish to modify the compiler or the microcode of the machine one is running on, or for that matter buy a processor which runs that calculation better. Similarly, it should be possible to define interfaces to swap management (note, for instance, the ITS paging parameters PagAhd and PagRange which warn the system of linear sweeps) which either define a particular regime of swapping or even allow swapping to be fully controlled by the user process (e.g. in addition to requesting swap-in of the faulted page, request several 'related' pages): a page fault becomes essentially a kind of subroutine invocation. Demand paging with its various refinements is a sound general-purpose method, but certainly other methods are possible when the application warrants. Stavros Macrakis ------------------------------ Date: 14 Jan 1982 1102-PST Subject: Re: big memory chips From: BILLW at SRI-KL I believe the largest RAM chip to date is an IBM 288K dynamic (in a *9 organization). There were rumors in Electronic engineering Times of an entirely new nonvolitile RAM technology that is alledged to be able to put 4M bits on a chip the size of todays ram chips using ordinary processes. Sort of two 5 volt chips with electron beams going inbetween them, if I recall (this was a couple of months ago). Most of the secrets are being kept under wraps until "they can be adaquatly protected". Bill W ------------------------------ From: William "Chops" Westfield Subject: more on big memories zurich -(dj)--nippon electric co. (nec) of japan expects to post higher profits for the year ending march 31, senior executive vice president m. hirota told zurich bankers thursday. : : hirota said nec has -solved all the technical complications for mass production of a 256 kilobit random access memory (kram) circuit, which would quadruple the memory capacity of computers and telecommunications equipment which currently use 64 kram circuits. demand for the 64 kram circuits is still growing and should peak in 1984 or 1985, he predicted, adding that demand for 256 kram circuits should become significant by 1986. : : ------------------------------ End of WorkS Digest ******************* ------- ----------------------------------------------------------------- gopher://quux.org/ conversion by John Goerzen of http://communication.ucsd.edu/A-News/ This Usenet Oldnews Archive article may be copied and distributed freely, provided: 1. There is no money collected for the text(s) of the articles. 2. The following notice remains appended to each copy: The Usenet Oldnews Archive: Compilation Copyright (C) 1981, 1996 Bruce Jones, Henry Spencer, David Wiseman.