<div dir="ltr">Hi David,<div><br></div><div>This is wonderful data and demonstrates the viability of this feature. I think this has alleviated the concerns regarding file locking.</div><div><br></div><div>As far as the implementation of the feature, I think we will probably want the following incremental steps:</div><div>a) implement the core merging logic and add to buffer API a primitive for merging two buffers</div><div>b) implement the file system glue to extend this to the filesystem API's (write_file etc.)</div><div>c) implement a profile filename format string which generates a random number mod a specified amount (strawman: `LLVM_PROFILE_FILE=default.profraw.%7u` which generates a _u_nique number mod 7. Of course, in general it is `%<N>u`)</div><div><br></div><div><div> b) depends on a), but c) can be done in parallel with both.</div></div><div><br></div><div>Does this seem feasible?</div><div><br></div><div>-- Sean Silva</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Mar 1, 2016 at 2:55 PM, Xinliang David Li <span dir="ltr"><<a href="mailto:davidxl@google.com" target="_blank">davidxl@google.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">I have implemented the profile pool idea from Mehdi, and collected performance data related to profile merging and file locking. The following is the experiment setup:<div><br></div><div>1) the machine has 32 logical cores (Intel sandybridge machine/64G memory)</div><div>2) the workload is clang self build (~3.3K files to be built), and the instrumented binary is Clang.</div><div>3) ninja parallelism j32</div><div><br></div><div>File systems tested (on linux)</div><div>1) a local file system on a SSD drive<br></div><div>2) tmpfs</div><div><div>3) a local file system on a hard disk</div></div><div>4) an internal distributed file system</div><div><br></div><div>Configurations tested:</div><div>1) all processes dump to the same profile data file without locking (this configuration of course produces useless profile data in the end, but it serves as the performance baseline)</div><div>2) profile-merging enabled with pool sizes : 1, 2, 3, 4, 5, 10, and 32</div><div>3) using LLVM_PROFILE_FILE=..._%p to enable each process to dump its own copy of profile data (resulting in ~3.2K profile data files in the end). This configuration is only tested on some FS due to size/quota constraints.</div><div><div><br></div><div>Here is a very high level summary of the experiment result. The longer writing latency it is, the more file locking contention is (which is not surprising). In some cases, file lock has close to zero overhead, while for FS with high write latencies, file locking can affect performance negatively. In such cases, using a small pool of profile files can completely recover the performance. The size of the required pool size is capped at a small value (which depends on many different factors: write latency, the rate at which the instrumented binary retires, io throughput/network bandwidth etc).</div><div><br></div><div>1) SSD</div><div><br></div><div>The performance is almost identical across *ALL* the test configurations. The real time needed to complete the full self build is ~13m10s. There is no visible file contention with file locking enabled even with pool size == 1.</div><div><br></div><div>2) tmpfs</div><div><br></div><div>only tested with the following configs</div><div>a) shared profile with no merge</div><div>b) with merge (pool == 1), with merge (pool == 2)</div><div><br></div><div>Not surprisingly, the result is similar to SSD case -- consistently finished building in a little more than 13m.</div><div><br></div><div>3) HDD</div><div><br></div><div>With this configuration, file locking start to show some impact -- the write is slow enough to introduce contention.</div><div><br></div><div>a) Shared profile without merging: ~13m10s</div><div>b) with merging</div><div> b.1) pool size == 1: ~18m20s</div><div> b.2) pool size == 2: ~16m30s</div><div> b.3) pool size == 3: ~15m55s</div><div> b.4) pool size == 4: ~16m20s</div><div> b.5) pool size == 5: ~16m42s</div><div>c) >3000 profile file without merging (%p) : ~16m50s</div><div><br></div></div><div>Increasing the size of merge pool increases dumping parallelism -- the performance improves initially but when it is above 4, it starts to degrade gradually. When the HDD IO throughput is saturated at that point and increasing parallelism does not help any more.</div><div><br></div><div>In short, with profile merging, we just need to dump 3 profile files to achieve the same build performance that dumps >3000 files (the current default behavior).</div><div><br></div><div>4) An internal file system using network attached storage</div><div><br></div><div>In such a file system, the file write has relatively long latency compared with local file systems. The backend storage server does dynamic load balancing so that it can achieve very high IO throughput with high parallelism (at both FE/client side and backend). </div><div><br></div><div>a) Single profile without profile merging : ~60m</div><div>b) Profile merging enabled:</div><div> b.1) pool size == 1: ~80m</div><div> b.2) pool size == 2: ~47m</div><div> b.3) pool size == 3: ~43m</div><div> b.4) pool size == 4: ~40m40s</div><div> b.5) pool size == 5: ~38m50s</div><div> b.6) pool size == 10: ~36m48s</div><div> b.7) pool size == 32: ~36m24s</div><div>c) >3000 profile file without profile merging (%p): ~35m24s</div><div><br></div><div>b.6), b.7) and c) have the best performance among all.</div><div><br></div><div>Unlike in HDD case, a) has poor performance here -- due to low parallelism in the storage backend. </div><div><br></div><div>With file dumping parallelism, the performance flats out when the pool size >= 10. This is because the client (ninja+clang) system has reached its peak and becomes the new performance bottleneck.</div><div><br></div><div>Again, with profile merging, we only need 10 profile data file to achieve the same performance as the default behavior that requires >3000 files to be dumped.</div><div><br></div><div>thanks,</div><div><br></div><div>David</div><div><br></div><div><br></div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote"><span class="">On Sun, Feb 28, 2016 at 12:13 AM, Xinliang David Li <span dir="ltr"><<a href="mailto:davidxl@google.com" target="_blank">davidxl@google.com</a>></span> wrote:<br></span><div><div class="h5"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><br><div class="gmail_quote"><span>On Sat, Feb 27, 2016 at 6:50 PM, Sean Silva <span dir="ltr"><<a href="mailto:chisophugis@gmail.com" target="_blank">chisophugis@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr">I have thought about this issue too, in the context of games. We may want to turn profiling only for certain frames (essentially, this is many small profile runs).<div><br></div><div>However, I have not seen it demonstrated that this kind of refined data collection will actually improve PGO results in practice.</div><div>The evidence I do have though is that IIRC Apple have found that almost all of the benefits of PGO for the Clang binary can be gotten with a handful of training runs of Clang. Are your findings different?</div></div></blockquote><div><br></div></span><div>We have a very wide customer base so we can not claim one use model is sufficient for all users. For instance, we have users using fine grained profile dumping control (programatically) as you described above. There are also other possible use cases such as dump profiles for different periodical phases into files associated with phases. Later different phase's profile data can be merged with different weights.</div><span><div> <br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div><br></div><div>Also, in general, I am very wary of file locking. This can cause huge amounts of slowdown for a build and has potential portability problems. </div></div></blockquote><div><br></div></span><div>I don't see much slow down with a clang build using instrumented clang as the build compiler. With file locking and profile merging enabled, the build time on my local machine looks like:</div><div><br></div><div><div>real 18m22.737s</div><div>user 293m18.924s</div><div>sys 9m55.532s</div></div><div><br></div><div>If profile merging/locking is disabled (i.e, let the profile dumper to clobber/write over each other), the real time is about 14m. </div><span><div> <br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div>I don't see it as a substantially better solution than wrapping clang in a script that runs clang and then just calls llvm-profdata to do the merging. Running llvm-profdata is cheap compared to doing locking in a highly parallel situation like a build.</div></div></blockquote><div><br></div></span><div>That would require synchronization for merging too. </div><div><br></div><div>From Justin's email, it looks like there is a key point I have not made clear: the on-line profile merge is a very simple raw profile to raw profile merging which is super fast. The end result of the profile run is still in raw format. The raw to indexed merging is still needed -- but instead of merging thousands of raw profiles which can be very slow, with this model, only one raw profile input is needed.</div><div><br></div><div>thanks,</div><div><br></div><div>David</div><span><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div><br></div><div><div><br></div><div>-- Sean Silva</div></div></div><div class="gmail_extra"><br><div class="gmail_quote"><div><div>On Sat, Feb 27, 2016 at 6:02 PM, Xinliang David Li via llvm-dev <span dir="ltr"><<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>></span> wrote:<br></div></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div><div><div dir="ltr">One of the main missing features in Clang/LLVM profile runtime is the lack of support for online/in-process profile merging support. Profile data collected for different workloads for the same executable binary need to be collected and merged later by the offline post-processing tool. This limitation makes it hard to handle cases where the instrumented binary needs to be run with large number of small workloads, possibly in parallel. For instance, to do PGO for clang, we may choose to build a large project with the instrumented Clang binary. This is because<div> 1) to avoid profile from different runs from overriding others, %p substitution needs to be specified in either the command line or an environment variable so that different process can dump profile data into its own file named using pid. This will create huge requirement on the disk storage. For instance, clang's raw profile size is typically 80M -- if the instrumented clang is used to build a medium to large size project (such as clang itself), profile data can easily use up hundreds of Gig bytes of local storage. </div><div>2) pid can also be recycled. This means that some of the profile data may be overridden without being noticed.</div><div><br></div><div>The way to solve this problem is to allow profile data to be merged in process. I have a prototype implementation and plan to send it out for review soon after some clean ups. By default, the profiling merging is off and it can be turned on with an user option or via an environment variable. The following summarizes the issues involved in adding this feature:</div><div> 1. the target platform needs to have file locking support</div><div> 2. there needs an efficient way to identify the profile data and associate it with the binary using binary/profdata signature;</div><div> 3. Currently without merging, profile data from shared libraries (including dlopen/dlcose ones) are concatenated into the primary profile file. This can complicate matters, as the merger also needs to find the matching shared libs, and the merger also needs to avoid unnecessary data movement/copy;</div><div> 4. value profile data is variable in length even for the same binary.</div><div><br></div><div>All the above issues are resolved and clang self build with instrumented binary passes (with both j1 and high parallelism). </div><div><br></div><div>If you have any concerns, please let me know.</div><div><br></div><div>thanks,</div><div><br></div><div>David</div><div><br></div></div>
<br></div></div>_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
<br></blockquote></div><br></div>
</blockquote></span></div><br></div></div>
</blockquote></div></div></div><br></div>
</blockquote></div><br></div>