I had previously send an E-Mail to the Mesos DEV list on Tue 2/14/2017 11:29
AM. I've attached that message here for reference (hopefully the attachment
will make it across).
Work has proceeded, and I'm working out some commits to apply to the changes
for Windows (cmake only).
While I was asked to work out timing for Linux platforms, I was unable to do so
as the Linux code doesn't currently build with precompiled headers due to
overloading of the std namespace (Set/set, as I recall). We could hack sources
to build with precompiled headers, but this would be throw-away work and thus
doesn't seem all that useful an engineering expenditure. But without the
permanent changes needed there to avoid namespace pollution, or hack/throw away
changes, it's not possible to get timing tests for PCH use on Linux at this
time.
For Windows, here are the results of timing tests that I've done:
Build times on Windows (no PCH):
1. Full build time (everything including 3rd party products): 24:49.47
(1489.47s total)
2. Time to rebuild all of Mesos itself: 21:8.63 (1268.63s total)
3. Time for a build with no changes: 0:30.47 (30.47s total)
4. Modified file 3rdparty/stout/include/stout/os/os.hpp (added a comment)
5. Time for an incremental build: 36:55.55 (2215.55s total)
(Very odd that an incremental takes longer than a full build!)
6. Time for an incremental build (just mesos-agent): 0:14:24.73 (864.73s
total)
7. Time for an incremental build, mesos-agent only: 0:16:42.57 (1002.57s
total)
(Includes speedup to linker flags)
Build times on Windows (with PCH):
1. Full build time (everything including 3rd party products): 19:54.49
(1194.49s total)
2. Time to rebuild all of Mesos itself: 0:15:49.58 (949.58s total)
3. Time for a build with no changes: 0:0:21.72 (21.72s total)
4. Modified file 3rdparty/stout/include/stout/os/os.hpp (added a comment)
5. Time for an incremental build: 0:34:40.64 (2080.64s total)
6. Time for an incremental build (just mesos-agent): 0:11:53.57 (713.57s
total)
7. Time for an incremental build, mesos-agent only: 0:10:43.10 (643.10s total)
(Includes speedup to linker flags)
To be clear here, more work can be done. We pretty much just optimized building
mesos-agent. There are other targets that would benefit from PCH as well.
However, unless/until we start doing major portions of work in those areas, it
may not make sense. (For example, we're unlikely to do major pieces of redesign
to most of our 3rd party components, nor do we have the ability to change those
build processes anyway.)
The above timing tests include the time to actually build the PCH. In the full
build case, the reduction is close to 25%. In the incremental build case (#7),
the reduction is closer to 33%. In both cases, we believe that this change is
worthwhile.
Any questions/concerns? If not, I'll work on finalizing the commits and getting
them tested in the next few days.
/Jeff
--- Begin Message ---
Proposal For Build Improvements
The Mesos build process is in dire need of some build infrastructure
improvements. These improvements will improve speed and ease of work in
particular components, and dramatically improve overall build time, especially
in the Windows environment, but likely in the Linux environment as well.
Background:
It is currently recommended to use the ccache project with the Mesos build
process. This makes the Linux build process more tolerable in terms of speed,
but unfortunately such software is not available on Windows. Ultimately,
though, the caching software is covering up two fundamental flaws in the
overall build process:
1. Lack of use of libraries
2. Lack of precompiled headers
By not allowing use of libraries, the overall build process is often much
longer, particularly when a lot of work is being done in a particular
component. If work is being done in a particular component, only that library
need be rebuilt (and then the overall image relinked). Currently, since there
is no such modularization, all source files must be considered at build time.
Interestingly enough, there is such modularization in the source code layout;
that modularization just isn't utilized at the compiler level.
Precompiled headers exist on both Windows and Linux. For Linux, you can refer
to
https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgcc.gnu.org%2Fonlinedocs%2Fgcc%2FPrecompiled-Headers.html&data=02%7C01%7CJeff.Coffler%40microsoft.com%7C063d6dbc04304614f82f08d4550fc22d%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636226973566412987&sdata=cacmJfK7ofxwx%2B663qCWVw3QzDdg61Eqps7ZV1Trxd0%3D&reserved=0.
Straight from the GNU CC documentation: "The time the compiler takes to
process these header files over and over again can account for nearly all of
the time required to build the project."
In my prior use of precompiled headers, each C or C++ file generally took about
4 seconds to compile. After switching to precompiled headers, the precompiled
header creation took about 4 seconds, but each C/C++ file now took about 200
milliseconds to compile. The overall build speed was thus dramatically reduced.
Scope of Changes:
These changes are only being proposed for the CMake system. Going forward, the
CMake system is the easiest way to maintain some level of portability between
the Linux and Windows platforms.
Details for Modularization:
For the modularization, the intent is to simply make each source directory of
files, if functionally separate, to be compiled into an archive (.a) file.
These archive files will then be linked together to form the actual
executables. These changes will primarily be in the CMake system, and should
have limited effect on any actual source code.
At a later date, if it makes sense, we can look at building shared library
(.so) files. However, this only makes the most sense if the code is truly
shared between different executable files. If that's not the case, then it
likely makes sense just to stick with .a files. Regardless, generation of .so
files is out of scope for this change.
Details for Precompiled Header Changes:
Precompiled headers will make use of stout (a very large header-only library)
essentially "free" from a compile-time overhead point of view. Basically,
precompiled headers will take a list of header files (including very long
header files, like "windows.h"), and generate the compiler memory structures
for their representation.
During precompiled header generation, these memory structures are flushed to
disk. Then, when components are built, the memory structures are reloaded from
disk, which is dramatically faster than actually parsing the tens of thousands
of lines of header files and building the memory structures.
For precompiled headers to be useful, a relatively "consistent" set of headers
must be included by all of the C/C++ files. So, for example, consider the
following C file:
#if defined(windows)
#include <windows.h>
#endif
#include <header-a>
#include <header-b>
#include <header-c>
< - Remainder of module - >
To make a precompiled header for this module, all of the #include files would
be included in a new file, mesos_common.h. The C file would then be changed as
follows:
#include "mesos_common.h"
< - Remainder of module - >
Structurally, the code is identical, and need not be built with precompiled
headers. However, use of precompiled headers will make file compilation
dramatically faster.
Note that other include files can be included after the precompiled header if
appropriate. For example, the following is valid:
#include "mesos_common.h"
#inclue <header-d>
< - Remainder of module - >
For efficiency purposes, if a header file is included by 50% or more of the
source files, it should be included in the precompiled header. If a header is
included in fewer than 50% of the source files, then it can be separately
included (and thus would not benefit from precompiled headers). Note that this
is a guideline; even if a header is used by less than 50% of source files, if
it's very large, we still may decide to throw it in the precompiled header.
Note that, for use of precompiled headers, there will be a great deal of code
churn (almost exclusively in the #include list of source files). This will mean
that there will be a lot of code merges, but ultimately no "code logic"
changes. If merges are not done in a timely fashion, this can easily result in
needless hand merging of changes. Due to these issues, we will need a dedicated
sheppard that will integrate the patches quickly. This kind of work is easily
invalidated when the include list is changed by another developer,
necessitating us to redo the patch. [Note that Joseph has stepped up to the
plate for this, thanks Joseph!]
This is the end of my proposal, feedback would be appreciated.
--- End Message ---