Hello,
I am working on improving Mozilla startup times. It turns out that page 
faults(caused by lack of cooperation between user/kernelspace) are the 
main cause of slow startup. I need some insights from someone who 
understands linux vm behavior.

Current Situation:
The dynamic linker mmap()s  executable and data sections of our 
executable but it doesn't call madvise().
By default page faults trigger 131072byte reads. To make matters worse, 
the compile-time linker + gcc lay out code in a manner that does not 
correspond to how the resulting executable will be executed(ie the 
layout is basically random). This means that during startup 15-40mb 
binaries are read in basically random fashion. Even if one orders the 
binary optimally, throughput is still suboptimal due to the puny readahead.

IO Hints:
Fortunately when one specifies madvise(WILLNEED) pagefaults trigger 2mb 
reads and a binary that tends to take 110 page faults(ie program stops 
execution and waits for disk) can be reduced down to 6. This has the 
potential to double application startup of large apps without any clear 
downsides. Suse ships their glibc with a dynamic linker patch to 
fadvise() dynamic libraries(not sure why they switched from doing 
madvise before).

I filed a glibc bug about this at 
http://sourceware.org/bugzilla/show_bug.cgi?id=11431 . Uli commented 
with his concern about wasting memory resources. What is the impact of 
madvise(WILLNEED) or the fadvise equivalent on systems under memory 
pressure? Does the kernel simply start ignoring these hints?

Also, once an application is started is it reasonable to keep it 
madvise(WILLNEED)ed or should the madvise flags be reset?

Perhaps the kernel could monitor the page-in patterns to increase the 
readahead sizes? This may already happen, I've noticed that a handful of 
pagefaults trigger > 131072bytes of IO, perhaps this just needs tweaking.

Thanks,
Taras Glek

PS. For more details on this issue see my blog at ...
[ message continues ]

Try tuning /sys/block/<dev>/queue/read_ahead_kb and see if that
makes any difference - that's the default maximum readahead for the
given block device and defaults to 128k.

There has been some recent work to increase the default readahead
size, so if changing the default improves performance then perhaps
a fix for your problem is already in the works?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com
--

[ message continues ]

Almost certainly teaching my grandmother to suck eggs, but are you aware
of the work Michael Meeks has done on improving openoffice.org startup time?
-- 
Roland Dreier <rolandd@cisco.com> || For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/index.html
--

[ message continues ]

Yes. There were some stones left unturned in the cold startup area. 
Turns out that every single large application suffers from low io 
throughput likely due to lack of cooperation between the dynamic linker 
and the kernel.
There is a glibc bug filed on that.

http://sourceware.org/bugzilla/show_bug.cgi?id=11431

Unfortunately, few userspace people seem to know exactly how madvise() 
hints behave, so I was hoping someone on LKML would clue me in.

Taras
--

[ message continues ]

It will throttle based on memory pressure.  In idle situations it will
eat your file cache, however, to satisfy the request.

Now, the file cache should be much bigger than the amount of unneeded
pages you prefault with the hint over the whole library, so I guess the
benefit of prefaulting the right pages outweighs the downside of evicting
some cache for unused library pages.

Still, it's a workaround for deficits in the demand-paging/readahead

It's a one-time operation that starts immediate readahead, no permanent

--

[ message continues ]

Define idle situations. Do you mean that madv(willneed) will aggresively 
readahead, but only while cpu(or disk?) is idle?
I am trying to optimize application startup which means that the cpu is 
I may be measuring this wrong, but in my experience the only change 
madvise(willneed) does in increase the length parameter to 
__do_page_cache_readahead(). My script is at 
http://hg.mozilla.org/users/tglek_mozilla.com/startup/file/6453ad2a7906/kernelio.stp 
.


Taras
--

[ message continues ]

Sorry.  I meant without memory pressure.  It will trigger readahead for the
whole page range immediately, unless the sum of free pages and file cache
pages is less than that.

So yes, it will be aggressive against the cache but should not touch things

Whether the page is read on a major fault or by means of WILLNEED,
they both end up calling this function.  It's just that faulting
does all the heuristics and WILLNEED will just force reading the
pages in the specified range.

But your question whether it would be reasonable to keep the region
WILLNEED madvised makes no sense.  It's just a request to prepopulate
the page cache from disk data immediately instead of waiting for
faults to trigger the reads.
--

[ message continues ]

Ok. Thanks for clarifying that. I was misinterpreting my io log.
Is there a way to force page faults from a particular memory mapping to 
do more readahead? Ie if WILLNEED is not used.


Have heuristics that read backwards been considered? Ie currently if one 
faults in page at offset 4096, that page a few pages following that will 
be preread. Would be interesting to try to preread pages before and 
after the page being faulted in.
For a graph of "backwards" io see the "Post-linker Fail" section in
http://blog.mozilla.com/tglek/2010/03/24/linux-why-loading-binaries-from-disk-sucks/


Taras
--

[ message continues ]

Hi Taras,


How about improve Fedora (and other distros) to preload Mozilla (and
other apps the user run at the previous boot) with fadvise() at boot
time? This sounds like the most reasonable option.

As for the kernel readahead, I have a patchset to increase default
mmap read-around size from 128kb to 512kb (except for small memory

This is interesting. I wonder how SuSE implements the policy.
Do you have the patch or some strace output that demonstrates the

Program page faults are inherently random, so the straightforward
solution would be to increase the mmap read-around size (for desktops
with reasonable large memory), rather than to improve program layout


Thank you :)

Cheers,
Fengguang
--

[ message continues ]

That's a slightly different usecase. I'd rather have all large apps 
startup as efficiently as possible without any hacks. Though until we 
Yes. Is the current readahead really doing read-around(ie does it read 
pages before the one being faulted)? From what I've seen, having the 
dynamic linker read binary sections backwards causes faults.
glibc-2.3.90-ld.so-madvise.diff in 
http://www.rpmseek.com/rpm/glibc-2.4-31.12.3.src.html?hl=com&cba=0:G:0:3732595:0:1... 


Program page faults may exhibit random behavior once they've started.

During startup page-in pattern of over-engineered OO applications is 
very predictable. Programs are laid out based on compilation units, 
which have no relation to how they are executed. Another problem is that 
any large old application will have lots of code that is either rarely 
executed or completely dead. Random sprinkling of live code among mostly 
unneeded code is a problem.
I'm able to reduce startup pagefaults by 2.5x and mem usage by a few MB 
with proper binary layout. Even if one lays out a program wrongly, the 
worst-case pagein pattern will be pretty similar to what it is by default.

But yes, I completely agree that it would be awesome to increase the 
readahead size proportionally to available memory. It's a little silly 
to be reading tens of megabytes in 128kb increments :)  You rock for 

Cheers,
Taras
--

[ message continues ]

Hi, Wu and Taras.

I have been watched at this thread.
That's because I had a experience on reducing startup latency of application
in embedded system.

I think sometime increasing of readahead size wouldn't good in embedded.
Many of embedded system has nand as storage and compression file system.
About nand, as you know, random read effect isn't rather big than hdd.
About compression file system, as one has a big compression,
it would make startup late(big block read and decompression).
We had to disable readahead of code page with kernel hacking.
And it would make application slow as time goes by.
But at that time we thought latency is more important than performance
on our application.

Of course, it is different whenever what is file system and
compression ratio we use .
So I think increasing of readahead size might always be not good.

Please, consider embedded system when you have a plan to tweak
readahead, too. :)

-- 
Kind regards,
Minchan Kim
--

[ message continues ]

Minchan, glad to know that you have experiences on embedded Linux.

While increasing the general readahead size from 128kb to 512kb, I
also added a limit for mmap read-around: if system memory size is less
than X MB, then limit read-around size to X KB. For example, do only
128KB read-around for a 128MB embedded box, and 32KB ra for 32MB box.

Do you think it a reasonable safety guard? Patch attached.

Thanks,
Fengguang

[ message continues ]

Thanks for reply, Wu.

I didn't have looked at the your attachment.
That's because it's not matter of memory size in my case.
It was alone application on system and it was first main application of system.
It means we had a enough memory.

I guess there are such many of embedded system.
At that time, although I could disable readahead totally with read_ahead_kb,
I didn't want it. That's because I don't want to disable readahead on
the file I/O
and data section of program. So at a loss, I hacked kernel to disable
readahead of
only code section.

-- 
Kind regards,
Minchan Kim
--

[ message continues ]

In general, the more memory size, the less we care about the possible

I would like to auto tune readahead size based on the device's
IO throughput and latency estimation, however that's not easy..

Other than that, if we can assert "this class of devices won't benefit
from large readahead", then we can do some static assignment.

Thanks,
Fengguang
--

[ message continues ]

A few month ago, I saw your patch about enhancing readahead.
At that time, many guys tested several size of USB and SSD which are
consist of nand device.
The result is good if we does readahead untile some crossover point.
So I think we need readahead about file I/O in non-rotation device, too.

But startup latency is important than file I/O performance in some machine.
With analysis at that time, code readahead of application affected slow startup.
In addition, during bootup, cache hit ratio was very small.

So I hoped we can disable readahead just only code section(ie, roughly
exec vma's filemap fault). :)

I don't want you to solve this problem right now.
Just let you understand embedded system's some problem



-- 
Kind regards,
Minchan Kim
--

[ message continues ]

Yeah, I've never heard of such a demand, definitely good to know it!

Thanks,
Fengguang
--

[ message continues ]

Boot time user space readahead can do better than kernel heuristic
readahead in several ways:

- it can collect better knowledge on which files/pages will be used
  which lead to high readahead hit ratio and less cache consumption

- it can submit readahead requests for many files in parallel,
  which enables queuing (elevator, NCQ etc.) optimizations



There are too many data in
http://people.mozilla.com/~tglek/startup/systemtap_graphs/ld_bug/report.txt

550 Can't open
/pub/linux/distributions/suse/pub/suse/update/10.1/rpm/src/glibc-2.4-31.12.3.src.rpm:
No such file or directory






Thank you. I guess the 128kb is more than ten years old..

Cheers,
--

[ message continues ]

The first part of the file lists sections in a file and their hex 
offset+size.

lines like 0 512 offset(#1) mean a read at position 0 of 512 bytes. 
Incidentally this first read is coming from vfs_read, so the log doesn't 
take account readahead (unlike the other reads caused by mmap page faults).

So
15310848 131072 offset(#2)=====================
eaa73c 1523c .bss
eaa73c 19d1e .comment

15142912 131072 offset(#3)=====================
e810d4 200 .dynamic
e812d4 470 .got
e81744 3b50 .got.plt
e852a0 2549c .data

Shows 2 reads where the dynamic linker first seeks to the end of the 
file(to zero out .bss, causing IO via COW) and the backtracks to
read in .dynamic. However you are right, all of the backtracking reads 
are over 64K.
Thanks for explaining that. I am guessing your change to boost 
Released it yesterday. Hopefully other bloated binaries will benefit 
from this too.

http://blog.mozilla.com/tglek/2010/04/07/icegrind-valgrind-plugin-for-optimizing-cold-...

Thanks a lot Wu, I feel I understand the kernel side of what's happening 
now.

Taras
--

[ message continues ]

Yes, every binary/library starts with this 512b read.  It is requested
by ld.so/ld-linux.so, and will trigger a 4-page readahead. This is not
good readahead. I wonder if ld.so can switch to mmap read for the
first read, in order to trigger a larger 128kb readahead. However this



It sounds painful to produce the valgrind log, fortunately the end
user won't suffer.

Is it viable to turn on the "-ffunction-sections -fdata-sections"
options distribution wide? If so, you may sell it to Fedora :)

Thanks,
Fengguang
--

[ message continues ]

Hi, Wu.


AFAIK, kernel reads first sector(ELF header and so one)  of binary in
case of binary.
in fs/exec.c,
prepare_binprm()
{
...
return kernel_read(bprm->file, 0, bprm->buf, BINPRM_BUF_SIZE);
}

But dynamic loader uses libc_read for reading of shared library's one.

So you may have a chance to increase readahead size on binary but hard on shared
library. Many of app have lots of shared library so the solution of
only binary isn't big about
performance. :(

-- 
Kind regards,
Minchan Kim
--

[ message continues ]

Correction with data: in my system, ld is doing one 832b initial read for every library:

        $ strace true
        execve("/bin/true", ["true"], [/* 44 vars */]) = 0
        brk(0)                                  = 0x608000
        mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7fb3b3ea0000
        access("/etc/ld.so.nohwcap", F_OK)      = -1 ENOENT (No such file or directory)
        mmap(NULL, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7fb3b3e9e000
        access("/etc/ld.so.preload", R_OK)      = -1 ENOENT (No such file or directory)
        open("/etc/ld.so.cache", O_RDONLY)      = 3
        fstat(3, {st_mode=S_IFREG|0644, st_size=140899, ...}) = 0
        mmap(NULL, 140899, PROT_READ, MAP_PRIVATE, 3, 0) = 0x7fb3b3e7b000
        close(3)                                = 0
        access("/etc/ld.so.nohwcap", F_OK)      = -1 ENOENT (No such file or directory)
        open("/lib/libc.so.6", O_RDONLY)        = 3
==>     read(3, "\177ELF\2\1\1\0\0\0\0\0\0\0\0\0\3\0>\0\1\0\0\0\320\353\1\0\0\0\0\0@"..., 832) = 832
        fstat(3, {st_mode=S_IFREG|0755, st_size=1379752, ...}) = 0
        mmap(NULL, 3487784, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7fb3b3931000
        mprotect(0x7fb3b3a7b000, 2097152, PROT_NONE) = 0
        mmap(0x7fb3b3c7b000, 20480, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x14a000) = 0x7fb3b3c7b000
        mmap(0x7fb3b3c80000, 18472, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x7fb3b3c80000
        close(3)                                = 0
        mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7fb3b3e7a000
        mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7fb3b3e79000
        arch_prctl(ARCH_SET_FS, 0x7fb3b3e796f0) = 0
        mprotect(0x7fb3b3c7b000, 16384, PROT_READ) = 0
        mprotect(0x7fb3b3ea1000, 4096, PROT_READ) = 0
        munmap(0x7fb3b3e7b000, 140899)          = ...
[ message continues ]

I have an older patch to create dynamic bitmaps based on the last 
run and only prefetch those pages. 

It wasn't entirely a win for everything and didn't work for shared
libraries, but with some additional tuning the approach still has
potential I think, by combining memory saving with prefetching.

ftp://firstfloor.org/pub/ak/pbitmap/INTRO
http://halobates.de/dp2.pdf

For your use case the algorithm would likely need some glibc support.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.
--

[ message continues ]

On Mon, 05 Apr 2010 15:43:02 -0700

Yes, the linker scrambles the executable's block ordering.

This just isn't an interesting case.  World-wide, the number of people
who compile their own web browser and execute it from the file which ld
produced is, umm, seven.

So I'd suggest that you always copy the executable to a temp file and
mv it back before running any timing tests.

--

[ message continues ]

Gentoo users? Linux From Scratch?

There are many more than 7 of us. Unless you are talking about the build 
environments always running some tool after ld which I am not aware of.


-- 
Zan Lynx
zlynx@acm.org

"Knowledge is Power.  Power Corrupts.  Study Hard.  Be Evil."
--

[ message continues ]

OK, eight then.

But I still don't think it's the case we should optimise for.  Not if
it impacts the common case even the slightest.  It'd be far far better
to change those distros to perform the very cheap, once-off step of
straightening out their executables (including shared libraries).

--

[ message continues ]

"make install" tends to copy. I am not aware of any Makefiles
that link directly to /usr/bin, and usually that wouldn't work
anyways because of permissions. copy fixes the problem.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.
--

[ message continues ]

... and those people who are executing the binary out of the build directory are probably running the regression  test (i.e., "make; make check") and on most developer machines, if they're lucky they have enough memory that the executable will still be in their page cache.   :-)

This being said, on modern file systems (i.e., btrfs, ext4, xfs, et. al), delayed allocation should mostly hide this problem; and if not, and the linker can estimate in advance how big the resulting binary will be, it could be modified to use the fallocate(2) system call.

-- Ted


--

[ message continues ]

... and those people who are executing the binary out of the build directory are probably running the regression  test (i.e., "make; make check") and on most developer machines, if they're lucky they have enough memory that the executable will still be in their page cache.   :-)

This being said, on modern file systems (i.e., btrfs, ext4, xfs, et. al), delayed allocation should mostly hide this problem; and if not, and the linker can estimate in advance how big the resulting binary will be, it could be modified to use the fallocate(2) system call.

-- Ted


--

[ message continues ]

I'm sorry that you don't find this interesting. I did not suggest that 
people compile their own browser to get a perfect layout. This is 
something that Mozilla can do when preparing builds and it's also 
something distributions can do. It just so happens that large parts of 
startup will be very similar for every single firefox install, might as 
You mean to get it into a cache or to hope to avoid fragmentation?  If 
you are suggesting this to avoid measuring the startup overhead of 
paging the binary in, I strongly disagee. It is the slowest part of 
firefox startup and needs to be addressed.

Taras


--

[ message continues ]

It's not a case we should optimise for.  It's perfectly reasonable for
the kernel to assume that the executable is reasonably well-laid-out on
disk.  And if is _isn't_ well-laid-out than that should be fixed in
userspace, because for simple locality-of-reference reasons, that's
always going to produce the fastest result.

Plus it's the common case as well - the executable was copied from DVD
or over the network or whatever.

Plus it's so utterly trivial for people who compile-their-own to
straighten the file out - just run cp!  These people have gone and
screwed up their file layout - they should fix that, rather than trying

No, nothing like that at all.

What I'm saying is that you shouldn't be testing or attempting to
optimise for files which were laid out by ld.  Because those files are
an utter mess - the block ordering is simply all over the place.  And
the great majority of people aren't using executables which were laid out
on disk by ld!

Instead, straighten out the block layout with `cp', then go and do the
testing and the optimisation.  Because if you're not taking this first
step then you're just not serious about performance at all!

Here's a small executable, as laid out by ld:

File offset	disk blocks
0-0: 		18383385-18383385 (1)
1-1: 		18383389-18383389 (1)
2-3: 		18383392-18383393 (2)
4-4: 		18383400-18383400 (1)
5-7: 		18383430-18383432 (3)
8-11: 		18383450-18383453 (4)
12-12: 		18383423-18383423 (1)
13-14: 		18383447-18383448 (2)
15-16: 		18383474-18383475 (2)
17-17: 		18383390-18383390 (1)
18-18: 		18383398-18383398 (1)
19-20: 		18383418-18383419 (2)
21-21: 		18383421-18383421 (1)
22-22: 		18383397-18383397 (1)
23-23: 		18383399-18383399 (1)
24-24: 		18383407-18383407 (1)
25-25: 		18383391-18383391 (1)
26-26: 		18383396-18383396 (1)
27-28: 		18383394-18383395 (2)
29-34: 		18383401-18383406 (6)
35-38: 		18383425-18383428 (4)
39-39: 		18383433-18383433 (1)
40-40: 		18383463-18383463 (1)
41-44: 		18383490-18383493 (4)
45-45: ...
[ message continues ]

Yeah ok. We are talking about different things. I meant the linker lays 
out the program badly, ie within the executable itself. Turns out that 
naively concatenating various compilation units makes for binaries that 
load slowly due to excessive seeking within the file. I wasn't talking 
about filesystem fragmentation. I agree that filesystem bustage caused 
by executing the linker isn't interesting.

Taras
--

[ message continues ]

My understanding was that this is usually gone when you use a delayed
allocation fs (xfs, ext4), unless your link sequence takes much longer
than the flush window.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.
--

[ message continues ]