December 2001 - Darwin-kernel

lmbench numbers
by rob＠terizla.org 03 Dec '01

03 Dec '01

nothing like an "objective" os comparison to liven up a relatively young list :> lmbench is a suite of microbenchmarks designed to measure various os functions. it's pretty well-accepted as a good way to profile an operating system. ( more info: http://www.bitmover.com/lmbench ) i see from the archives that lmbench numbers were last posted in may. hope i'm not revisiting the topic too soon. These numbers came from a g4 400 powerbook, with netbsd, linux, and os x installed. For kicks, you can see how results vary with several recent OS X updates. netbsd-powerpc is basically the GENERIC kernel, recompiled to not have lots of drivers for hardware not present on the laptop. linux is from the "benh" tree: it's got better support for the hardware, but as far as core os features should be similar to a kernel.org kernel. I've got 2 or three runs for each kernel, just in case there's an outlying freak result for one. An empty slot for a test means that for some reason the test was not run. (in the case of netbsd, i disabled the rpc tests because they just hung and never completed. the other no-results are mysteries to me) i would classify the load going on while these tests were running as "normal idle workstation" ( e.g. mozilla was loaded, but i wasn't actively browsing the web ) Darwin, or more likely the mach sublayer, has significantly higher overhead for several classes of operating system functions ( most notably, context switch, file latency, and VM latency ). would the darwin kernel hackers mind elaborating on some of the reasons why? Thanks ==rob -- Rob Latham Allentown, PA USA EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897 -- Rob Latham Allentown, PA USA EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897 Subject: lmbench numbers lmbench is a suite of microbenchmarks designed to measure various os functions. it's pretty well-accepted as a good way to profile an operating system. ( more info: http://www.bitmover.com/lmbench ) i see from the archives that lmbench numbers were last posted in may. hope i'm not revisiting the topic too soon. These numbers came from a g4 400 powerbook, with netbsd, linux, and os x installed. For kicks, you can see how results vary with several recent OS X updates. netbsd-powerpc is basically the GENERIC kernel, recompiled to not have lots of drivers for hardware not present on the laptop. linux is from the "benh" tree: it's got better support for the hardware, but as far as core os features should be similar to a kernel.org kernel. I've got 2 or three runs for each kernel, just in case there's an outlying freak result for one. An empty slot for a test means that for some reason the test was not run. (in the case of netbsd, i disabled the rpc tests because they just hung and never completed. the other no-results are mysteries to me) i would classify the load going on while these tests were running as "normal idle workstation" ( e.g. mozilla was loaded, but i wasn't actively browsing the web ) Darwin, or more likely the mach sublayer, has significantly higher overhead for several classes of operating system functions ( most notably, context switch, file latency, and VM latency ). would the darwin kernel hackers mind elaborating on some of the reasons why? Thanks ==rob -- Rob Latham Allentown, PA USA EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897

1 0

Re: Multiple OS's
by rob＠terizla.org 03 Dec '01

03 Dec '01

Eduardo I. Jim?nez (eijimenez(a)mac.com) said: > Hello again. > > I have another question. Is it possible to have Classic, OS X, Darwin > and Linux on the same machine?. I would really like to make some > comparisons among OS's to try to bring the best of them to Darwin. yep it's possible. you can add netbsd to the mix too. i've done it. and i'm just about to post the results :> ==rob -- Rob Latham Allentown, PA USA EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897

1 0

Multiple OS's
by Eduardo I. Jiménez 02 Dec '01

02 Dec '01

Hello again. I have another question. Is it possible to have Classic, OS X, Darwin and Linux on the same machine?. I would really like to make some comparisons among OS's to try to bring the best of them to Darwin. Particularly I'm concerned about: - Quality scheduling: Solaris has the most impressive scheduling system. Almost no load (except for "while(1) { fork(); }" executed without limits) can seriously affect its responsiveness. I would like to help bring such scheduling performance to Darwin. - I/O Performance: although things have improved dramatically, I think Darwin's I/O performance is not yet at best levels (CAUTION: I'm judging from OS X since I have not yet a Darwin copy, I may be wrong, it is up to you to prove me that), however I need to setup testing plans and tools, if you don't already have some. I would like to test I/O at the different arenas and draw conclusions. These arenas are basically network and disk I/O. USB and FireWire I/O seem to be fine, but testing and benchmarks should prove that too.

1 0

Re: Newbie
by Eduardo I. Jiménez 02 Dec '01

02 Dec '01

Now, what do the stack and datasize limits mean?. I used to think that in modern virtual memory systems there was no limit to a process datasize or stack space but the VM limit itself. What did I get wrong?. In Linux these values default to 8192KB and unlimited respectively. Any reason why OS X has 6144KB (odd amount, don't you think?) for datasize and 512KB for stack size. Sorry, stack size should be limited, but datasize shouldn't, or should it? Thanks Eduardo Jimenez

1 0

Newbie
by Eduardo I. Jiménez 02 Dec '01

02 Dec '01

Hello everyone. I am a newbie to kernel development. I am using MacOS X and I'm trying to get Darwin (it's hell not having DSL or Cable modem available in my area, do you think the Apple Store would burn it for me?). Anyway, I am pretty experienced with C/C++ and UNIX and still remember many OS basics from university classes in OS, distributed and parallel systems. I would like to know where to start. The first thing I would like to ask, before getting to any specific task is, do you have any performance testing plan or system benchmark that could allow us to test the performance of Darwin (in all areas), both to test performance of kernel improvements and to compare those results to other OSs or UNIXes?. I would like to focus on kernel performance (scheduling, memory management, drivers, multithreading, SMP). I have one question for which you may have an answer. Why is the maxproc limit in OS X set so low?. Granted, in most Linuxes its 256 (and that's the max) but in OS X 10.1.1 its 100. If one is to have an enterprise server, this may seem too little. Now, I don't know what Darwin's processing unit is, threads or processes. I guess it should be processes, which would mean that this limit does not apply to threads, which would mean that a process could have an (unlimited?) number of them. Am I right?. Now, what do the stack and datasize limits mean?. I used to think that in modern virtual memory systems there was no limit to a process datasize or stack space but the VM limit itself. What did I get wrong?. In Linux these values default to 8192KB and unlimited respectively. Any reason why OS X has 6144KB (odd amount, don't you think?) for datasize and 512KB for stack size. Another question, is there any documentation that would describe the internals of the kernel so I can start looking at the source code? Thanks a lot, and I hope I can help improve what I think is a remarkable effort. Eduardo Jimenez

1 0