Messages in this thread | | | Date | Wed, 18 Mar 1998 16:42:39 +0100 (CET) | From | Andreas Kostyrka <> | Subject | Re: some memory/swap thoughts |
| |
On Wed, 18 Mar 1998, Jan Gyselinck wrote:
> Now, 5 meg as a cache, isn't that too much?? Some people think this is Nope, I routinly run with 80MB swap (here and at home.) And it's a difficult question, because there are tradeoffs: Either you want one bzip to run fast, or you want a dozen bzip par. run with high troughput. (That's nicely demonstrated by EIDE/SCSI on a multiuser server: Replacing a 5 yrs old SCSI disc with a new EIDE that was in benchmarks 3-4 times faster resulted in grinding teh server to a halt. I had to replace EIDE disc with a new SCSI disc, which again in benchmarks is about 50-60% of the speed of the EIDE disc, but I can do background kernel recompiles without anyone even noticing :) ).
> needed, well I can tell you, it isn't. A year ago, I did some testing on > a 486DX33 with 8 meg, running DOS/winslows 3.11. I tried different > cache-sizes (with pc-cache) and measured the speedup while starting up Comparing the Linux cache subsystem to SMARTDRV is like doing performance testing on LADA and then applying the findings to BMW/Cadillac/... They may hald true, but they may hold also not true! > MS-Word. Maybe you think that's not the way to test this, but why not? > You test the speed-up in real-life applications, because that's what you > do all day. So, speedup from 0 to 64 kB cache, 20 seconds, from 64 to 128 > kB cache, 14 seconds, from 128 to 256 kB, 6 seconds, from 256 to 512 kB, 3 > seconds, and from 512 to 1024 kB, 1 second. Now why in gods name would Again, on a Single-User box a different memory management may seem better. (But what happens if the single-user becomes a power user and begins to ``multitask'' heavily?) > one want 5 meg of cache? It will increase the speed of disk-activity with > maybe 1 second of a 2,5 meg cache. My oppinion is that for 16 meg of > memory, the minimum-limit for a disk-cache should be 256kB, not 5 meg!! Nope, that sucks. I know how NeXT with 800kb-2MB cache runs on the same hardware as Linux. Limiting the cache is not a very clever thing. > > Okay, you say, but what about all those idling programs that are stuck in > memory, and just take up memory from the cache. I know, I know, there are > cases where it's needed to run such programs, but not always. People are > running to many idle programs these days! Why do you think there is a Nope. By definitation a process that does a sleep(3600); should not be a hurden on the system. -> That's usual POSIX coding understanding. And not all can be handled by inetd, and not all should be handled. > inet-daemon? So that there don't need to be a dozen idle processes who > are checking if there isn't something knocking on there port. Running 6 They are not checking. They are blocked. And that means they are eligible for swapping. > or more agetty's? There exists something like a console spawn daemon, you Nice. And what if each of the ``getty's'' does a different thing? (One menu for system halt/restart/etc., one menu to restart gpm [Did you notice it's quite impossible to find GOOD 3 button trackballs today, especially if you don't want to pay MEGABUCKs? I've got one here, that seems to ``powerdown'' and needs to be restarted with button pressed, ...]
> If I run a bash on a console, and I do something on another console for a > while, and I return to the first one, I want the bash-process to respond > immediatly to my key-strokes. I don't like to wait for it until it's > loaded from swap. My opinion is this: if I run something, it is because > it needs to run, and it must be able to respond immediatly. If this is But Linux is Unix and by definition is a server OS. If you want a buggy scheduler implementation, than go buy OS/2 or NT *g*. (I've seen a nice SQL server for OS/2, but the scheduler was quite broken: It automatically decreased the priority of the server if you minimized it. So the Server had to run in foreground all the time, ...) > not so, I wont run it. I don't have memory to throw around, I need every > bit. (And no, memory for a portable is not that cheap) Than you want to use it more cleverly than to have say 900kb wasted on bash blocked on read(0,..,..);
> So think about this, when you people change something in the > memory-management of linux, because it'll run on low-budget and > high-budget systems... The point is, that Linux is as a server OS more optimized for the multitasking/multiuser troughput. And that's the complicated one: If you want to run one thing a time, than DOS quite cut it :)
> Jan Gyselink > for the moment a swapping linux-user Nothing bad about swapping. (And running bzip on a underpowered machine was always painful. If it wouldn't be so painful, than bzip would have long replaced gzip, right?)
> PS: i hope .90 is better, 'll try it tonight, but after I rebooted, > because my console is messed up by Xwindows, who didn't restore the state > after it finished (maybe caused by to much swapping?????) Why should swapping influence this? *wonder*
Andreas
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu
| |