From: bmc@kiowa.eng.sun.com (Bryan Cantrill)
Subject: Re: But why?
Date: 1996/10/29
Message-ID: <5543dq$qjd@engnews2.Eng.Sun.COM>
references: <54990q$n5e@caip.rutgers.edu>
organization: Brown University
keywords: none
newsgroups: comp.sys.sun.hardware



In article <54990q$n5e@caip.rutgers.edu>,
David S. Miller  wrote:
>bacon@mtu.edu (Jeff Bacon) writes:
>
>Since I have been able to find an intelligent posting in this thread,
>I will respond to it and explain what I can as chief architect of the
>SparcLinux port.
>
>> of course, then, the obvious question that comes up is WHY is it that
>> solaris has such higher overhead costs in doing things? 
>> 
>> obviously there's more code to work through to do any given thing. 
>> someone must have thought it necessary. but...why? 
>> 
>> obviously it's got lots of extra crud from SVR4. why not pitch it? 
>> 
>
>The answer to this is pretty straight forward actually.  The main
>points of interest are:
>
>1) Solaris's networking stack, in all of it's incantations (one breed
>   of it was the Lochman code in 2.0, 2.1 and early 2.2 releases, then
>   it was rewritten by another company for 2.3 onward) is SVR4 streams
>   based.  The performance penalty, even with lots of tricks, for
>   using a SVR4 streams networking architecture is well known.
>   Someone who happens to have a 2.2 Solaris CD around, or even a 2.3
>   Solaris CD, should install that thing and run lmbench on it to see
>   what "pure Streams based networking" without the tricks can really
>   do.
>
>   Linux on the other hand has a "no bullshit" networking architecture
>   that is not streams based.  Yet we also take advantage of the many
>   known networking performance enhancements that exist in the
>   research realm (ie. copy/checksum, the van jacobson hacks, etc.)
>
>2) Linux is light weight, Solaris is a pig.
>
>   One of the most critical things that contributes to performance
>   is cache/tlb footprint of the operating system.  Linux being small
>   (yet still provide a full POSIX unix environment!) solves the cache
>   footprint problem in a big way.  I've solved the TLB footprint
>   using Linux's small size and a Sparc specific trick.
>
>   The MMU's present on the sun4m/sun4d line of Sun machines possess a
>   three level page table scheme.  Using this, one has the capability
>   to use the normal 4k sizes pages, and also larger 256k and 16MB
>   sized pages.  The average TLB on these machines has 32 or 64
>   entries to cache these pte's, if the entry is not in the TLB
>   hardware has to go out to the memory bus and walk the software page
>   tables to "reload" the TLB so that the translation can be
>   satisfied.
>
>   This "miss processing" is very expensive.  Under SunOS and Solaris,
>   they do not take advantage of the 16MB and 256k sized pages to map
>   the operating system.  Therefore those two systems take many misses
>   in the TLB during even the most rudimentary trap into the kernel.   
>   However under Linux the TLB misses for the OS are quite minimal.
>   In fact I will gave an example:
>
>      On your average SparcClassic with a 32 entry TLB, consider such
>      a machine with 24MB of memory installed.  Under Linux I can map
>      the entire operating system (sans IO device register mappings
>      and Lance Ethernet DMA) in 3 (count 'em, 3!) TLB entries.  These
>      3 entries are enough to allow the kernel to access an arbitrary
>      physical page from kernel space.
>
>      Under Solaris, the OS would need 3 + (24MB / 256K) + (24MB / 4K)
>      TLB entries to map this same amount of space.  For a great many
>      number of operations, it is quite easy for an OS with this page
>      table strategy to blow the entire user context out of the
>      hardware TLB.  Which in turn means more processor stalls (in
>      fact many) for both the user level processes and the operating
>      system.
>
>      Result?  Severe degradation in performance for the latter
>      scheme.
>
>3) Every BSD and SVR4 based system today, except for Linux, has a very
>   broken System call mechanism.
>
>   You'd think that when people put together function call conventions
>   for a particular processor, the OS people would take a look at this
>   and find a way to take advantage of this.  In fact, believe it or
>   not, they have not to this very day.
>
>   Linux from day one, takes advantage of the procedure call
>   conventions on a particular architecture so that it can process
>   system calls in the most expediant way possible.  I will give
>   an example on the Sparc to prove this:
>
>    Consider your average 3 argument system call.  The user level
>    code does something like this:
>
>	mov	%arg0, %o0
>	mov	%arg1, %o1
>	mov	%arg2, %o3
>	mov	SYSTEM_CALL_NUMBER, %g1
>	t	SYSCALL_TRAP
>
>    At this point control reaches the operating system, it must
>    prepare to handle this request from the user.  On the Sparc, this
>    is either a two step or a three step process depending upon
>    whether you are doing it in the traditional broken UNIX way or the
>    clean, fast, and superior Linux way.  First I will show the Linux
>    method for doing this:
>
>	1) Step one, jump onto the kernel stack for this task
>	   and make sure the kernel has a register window to
>	   operate in safely.
>
>	   For Linux the code path for this runs at ~18 instructions
>	   for the common case (the kernel already has a valid
>	   register to use so now saving needs to be done).  It runs
>	   at ~42 instructions for the second most common case (the
>	   kernels needs to allocate a new register window and the
>	   user has a valid stack pointer) and ~82 instructions for
>	   the least common case (kernel needs a window, user has
>	   an invalid stack pointer, thus the kernel needs to save
>	   the user's window into a special per-task save area).
>
>	2) Take the system call number, check for valid value, use
>	   this to offset into a table of system call function ptrs.
>	   Move arguments into place and perform the syscall.
>
>	   Basically this is a simple operations a looks something
>	   like:
>
>	sll	%g1, 2, %l4		! produce offset
>	ld	[%l7 + %g1], %l7	! syscall ptr base was in %l7
>	SAVE_ALL			! perform step #1 above
>	mov	%i0, %o0
>	mov	%i1, %o1
>	mov	%i2, %o2
>	mov	%i3, %o3
>	mov	%i4, %o4
>	jmp	%l7, %o7
>	 mov	%i5, %o5
>
>	   That is it, that is the entire system call under Linux.
>
>    Under Solaris/SunOS things are wildly different.  Step one is
>    basically the same, but step 2 is disgustingly inefficient for
>    those systems.  Basically they do:
>
>	2) Call common system_call() C function.
>
>	3) This routine allocates a "system call argument package"
>	   structure on the kernel stack.  This is wasteful because
>	   we already have all of this information in registers or
>	   in guarenteed save areas.
>
>	4) Then this routine determines the function to call, and
>	   passed this "package" of arguments to the routine.
>
>	5) Every system call which expects arguments then must
>	   "unpack" this structure to get at the copy of the arguments
>	   again highly inefficient.
>
>   For every system call the system performs, you eat this unnecessary
>   overhead under SunOS/Solaris, under Linux only the bare minimum is
>   performed to do the system call successfully.
>
>4) Solaris cannot even do it's own optimizations correctly because
>   SunPRO is a broken compiler.
>
>   I won't make such a statement without explaining this with real
>   facts, here goes.
>
>   A neat part of the Sparc ABI is that it leaves you with a few
>   processor registers that the C compiler is not allowed to use
>   in the code it produces.  Two of which are "%g6" and "%g7".
>   A problem in unix kernels is that you are constantly accessing
>   the current tasks control structure ('proc' and 'uarea' on
>   traditional UNIX's, the 'task_struct" under Linux).  Hey, why
>   not put these pieces of information in those "extra" registers
>   and avoid the address computation all the time?  Yes, very
>   brilliant idea.
>
>   Under Solaris the trap entry code places the uarea and proc ptr
>   in %g6 and %g7.  Under Linux the trap entry code places the
>   current processes task_struct in %g6.  Now here comes where the
>   implementations differ.
>
>   Under Solaris all of the so called "locore" (basically all the
>   gook which has to be written in raw assembly) code can directly
>   take advantage of this.  However, the C code cannot do this
>   because SunPRO lacks a way for you to tell the compiler that
>   "hey you don't need to load things, it's already in these
>    hard coded registers"  So they have the C code call these little
>   assembly stubs to get the values:
>
>	get_uarea:
>		retl
>		 mov	%g6, %o0
>
>	get_proc:
>		retl
>		 mov	%g7, %o0
>
>   That is gross, why even do the optimization in the first place?
>
>   Now GCC has a way to fully take advantage of such an optimization,
>   basically all I have to do is put the following in a header file.
>
>	register struct task_struct *current asm("g6");
>
>   Tada, now GCC will fully understand what I have done for it.
>   Under SparcLinux this optimization alone took away 115 instructions
>   in the scheduler sources, and it took ~50 instructions out of the
>   exit() handling, and it took ~65 instructions out of the fork()
>   handling.
>
>So my question always is, in matters such as these.  Who are these
>processor cycles for anyways, the kernel or the user?  Think about
>this when you consider how much overhead is being saved from one OS to
>another, and to what scale this is occurring.
>
>I hope that explains some of it, and gives people at least some sort
>of idea of the kinds of things that makes Linux scream on just about
>any hardware.  If people would like more explainations like the above,
>I'd be more than happy to chat with you via email about it or
>similar.  I love talking about performance issues on various
>processors and systems.
>
>Oh, and one thing that has not been mentioned yet in this thread (and
>yes NetBSD/OpenBSD both have this as well, good work guys).  That
>SparcLinux kernel that gets all of this incredible performance runs on
>both sun4c and sun4m machines.  Sun Engineers way back when scratched
>their heads for months and couldn't figure out a way to pull it off
>(you need a seperate kernel image depending upon whether you are
>running on a sun4m or a sun4c, for SunOS/Solaris).  And on top of that
>Linux obviously pulls it off efficiently.
>
>One final note.  When you have to deal with SunSOFT to report a bug,
>how "important" do you have (ie. Fortune 500?) to be and how big of a
>customer do you have to be (multi million dollar purchases?) to get
>direct access to Sun's Engineers at Sun Quentin?  With Linux, all you
>have to do is send me or one of the other SparcLinux hackers an email
>and we will attend to your bug in due time.  We have too much pride in
>our system to ignore you and not fix the bug.
>
>David S. Miller
>davem@caip.rutgers.edu
>

Have you ever kissed a girl?

	- Bryan

----------------------------------------------------------------------
Bryan Cantrill, Solaris Performance.   bmc@eng.sun.com  (415) 786-3652