Archive for February, 2013

Where did my memory go ? – A detective story

Tuesday, February 12th, 2013

Here, is yet another follow up post on EDS memory consumption. For the last few days I’ve been tracking where memory is spent in EDS and our benchmarking tools, and it was a very interesting experience.

And I’m not just saying that ! it was very trying and it’s still a bit of an unsolved mystery to me (so please feel free to step in with your theories on the unsolved parts !).

It all started when Michael asked me to explain the funny spikes in the memory usage graph presented in the previous post. The first thing I did was to produce a more “bumpy” graph by disabling the slice allocator, yielding what is in some ways a more accurate account of actual memory usage:

Memory usage measured for 12,800 contacts with G_SLICE=always-malloc

Memory usage measured for 12,800 contacts with G_SLICE=always-malloc

Interestingly, I say “in some ways” above because; one of the elements that we have to consider is memory fragmentation; memory management is generally more optimal and less fragmented when the slice allocator is active.

What we are looking at above is a left to right graph of overall memory usage; measured after each and every operation that we run on the addressbook. Each “dot” can be associated to one of the various latency tests that we run for each and every build of EDS (indicated in the legend).

First of all let’s demystify the “curious humps” which occur mostly to the “Custom Light” (light blue) benchmarks but are also noticeable in other benchmarks. These “humps” occur for four dots at a time, particularly when performing suffix searches on contact fields that are not stored in the summary SQLite tables for quick searches.

This phenomenon is partly attributable to the fact that all contacts in the addressbook need to be individually examined (and the vcards individually parsed) when the given contact field is not stored in the SQLite tables individually (or what we refer to in EDS terms as “the summary”). I’m not really very concerned by these “spikes”; obviously the memory is reclaimed later on, however it is curious that this happens specifically for suffix matching and not for prefix matching (presumably lot’s of extra string duplications and normalizations are needed for the case insensitive suffix matching routines).

Now that that’s out of the way, it leads us to some of the…

More interesting parts

I was at first not satisfied with only this explanation, sure, it kindof explains the “funny humps” in the benchmark progress but… by taking a closer look at what else is actually happening… I needed a better explanation.

The portions of the presented memory usage graphs that interest me more are the memory growth observable over the course of the first four dots, as well as the curious memory growth that also occurs at the very end of the benchmarks.

So what is happening in these stages ?

First of all, it’s positive news to know that the number of automatically generated vcards used for testing are already in memory before the benchmarks start at all, in the above graph that represents 12,800 vcards all in memory before the first benchmark is measured. And then…

  1. The addressbook is initialized and created, so at the point of measuring the very first dot, we have 12,800 vcards in memory and an initialized EBookClient on the client side and an addressbook counterpart (SQLite database created and active SQLite connection) in the server side memory
  2. Next, at the second dot we’ve created 12,800 EContact objects in memory… the 12,800 EContacts and 12,800 vcard strings remain in memory throughout the benchmark progress. This second dot is about 45MB higher on the scale than the first dot, so it’s pretty safe to say that 12,800 EContact objects cost roughly 45MB of resident memory which will not be reclaimed for the duration of the benchmark progress.
  3. The third dot is measured directly after adding all the contacts to the addressbook, here we start to see some divergence in memory usage; notice that this costs roughly 25MB extra for EBookClient based benchmarks, but only about 5MB for EBook based benchmarks. Being a bit naive, I overlooked this detail at the beginning of the investigation… one of the notable differences in the EBook apis is that it was lacking in batch commands. So the major difference here is that EBook tests add contacts one by one over D-Bus, while the EBookClient tests add contacts in batches of 3200 contacts at a time.
  4. This fourth dot, is after fetching all contacts at once from the addressbook. Here is where I became seriously alarmed. For normal clients, this shows an approximate 30MB growth in memory consumption. So where did my memory go ? A simple case of amnesia ?! Note though, that the Direct Read Access (red) benchmark hardly increases in memory for a fetch of 12,800 contacts, good show.

Naturally, feeling embarrassed about the consequences of the evil fourth dot… I frantically started my search for memory leaks… first I blamed the obscure nature of C++ code and it’s attempts to hide memory management behind smart pointers…I tried to pin it as a memory leak in the actual benchmarking code (after all, I did just lose 30MB of memory… it must have gone somewhere… right ?)… but after some tracing around, I found that those returned contacts, stored by smart pointers or not, were properly finalized and freed, leaving me with this uncomfortable mystery still on my hands.

While most of my memory leak hunt revolved around explaining the 30MB memory overhead incurred from dot 3 to dot 4, I should mention that the last memory jump was also suspicious. This last memory jump (which seems to vary between a 10MB to 25MB increase depending on the benchmark type) is incurred by deleting all contacts in the addressbook. So how about that ? I’ve just deleted all the contacts, and now I’m using MORE memory than before ?

The following day…

… I ran the benchmarks in loops, for some I’ll share below because this is how I eventually solved the mystery case, I also ran the benchmarks (server and client) under valgrind, ran some various test cases with the server and test cases running under valgrind. But the alleged memory leak was not to be tracked. Some testing of the benchmarks running in a loop seemed to indicate that there was some memory growth over time, not very much so, but enough to make me believe there must be some leak and be determined to find out.

Finally, today…

… I let my laptop chug along and loop the benchmarks (at least some of them) with a huge 12,800 contact count (that takes time), so let’s share those enlightening results here:

Memory usage while benchmarking 12,800 contacts in the first iteration

Memory usage benchmarking 12,800 contacts in the second iteration

Memory usage benchmarking 12,800 contacts in the second iteration

Memory usage benchmarking 12,800 contacts in the third iteration

Memory usage benchmarking 12,800 contacts in the third iteration

These results would be better viewed from left to right instead of one on top of the other, but you get the idea. Just consider that the last dot in the first chart happens directly before the first dot in the following chart, and so on.

So, after viewing this data… we can see that in the second and third graph, memory we presumed to be lost, is eventually returned to the system (in other words, it was indeed only a case of temporary amnesia, and not a more severe degrading case of alzheimer’s)… This is very reassuring, numerous runs with valgrind also show no real evidence of memory leakage, which is also reassuring evidence that our EDS is leak free.

But, that still doesn’t really explain…

Where is that memory actually going ?

At this point I can only give you my best guess, but all of the clues seem to point towards D-Bus traffic:

  • At the second “dot” where contacts are added to the addressbook, EBook APIs adding only a single contact at a time seems to cost much much less than using EBookClient apis and adding the contacts in batches of 3200 contacts at a time.
  • At the third “dot” where a brute “fetch all contacts” call is made to the addressbook, we can see a huge increase in memory consumption all except for when using Direct Read Access mode. So when fetching a list of 12,800 contacts not using D-Bus, we don’t suffer from memory loss.
  • In the last suspicious “dot”, where we delete all contacts from the addresssbook at once, all benchmark types seem to suffer significant memory loss. In this case the client is sending a list of 12,800 contact UIDs over D-Bus to the addressbook (in Direct Read Access as well, since deleting contacts is a write operation).

My best guess ? this is all due to zero-copy IPC transfers implemented by D-Bus.

In other words (if you’ve read up to this point you probably don’t need any explanation), instead of the sender writing chunks of data to a socket, and the receiver reading bytes from a socket; the sender is owning some shared memory which is accessed directly by the receiver.

This shared memory is probably managed by the D-Bus daemon itself, so it would make sense that the daemon not release the shared memory straight away but instead reserve some head room in the case that further transfers might reuse that memory.

So how come the fourth dot where a batch of 12,800 vcards are passed to the client, is not reused by the last dot where all contacts are deleted ? … Because, when contacts are fetched the shared memory owner would have to be the sender, which is the addressbook server. However when contacts are deleted, it is the EBookClient user process which sends a list of 12,800 UIDs, in this case the owner of the shared memory should be the other, client process.

I’ll probably need to pursue some extra verifications to be sure, but this best guess is very compelling to me at this time.

In conclusion, this was a really interesting exercise, which I don’t hope to repeat very often… but I did learn a few things and it did put some things into perspective. First and foremost; measuring memory usage, when compared to just tracking and plugging leaks, is quite another story… a lot more tricky and probably not an exact science.

If you’ve got this far, I hope you’ve enjoyed this detective story… I did enjoy it.

Amendments

It’s probably bad form but I’ll just add this here, my theory is obviously false. As I’ve been informed (already) that D-Bus does not implement any such zero-copy mechanisms with shared memory… so there is still a huge memory fluxuation, definitely related to D-Bus usage, which I can’t readily explain.

 

An exercise in memory consumption analysis

Monday, February 4th, 2013

Hi again.

This is a follow up on my recent post on features and improvements to the Evolution Data Server that we’ve been working on at Openismus. Note that the previous post explains what we’ve done in greater detail, some of this post might not make sense without reading the aforementioned post.

As I was asked to write a more complete report on how each of our patch sets effect memory consumption in EDS, I went ahead and ran some further comparisons. As usual, Mathias’ benchmarks saved the day (while the original benchmark suite only generates memory consumption comparisons for a single run of contacts, I was easily able to produce charts for each individual run and compare them separately).

Actually I had postponed this post since I was hoping to update our final patch set for Direct Read Access apis before reporting my findings. It seems however that currently EDS master is in a period of transition and so I’ll postpone the new patch submissions until some temporary regressions in EDS are fixed (the code which does work with EDS master is however available on the branch).

Memory Usage Report

In order to get a grasp of the impacts on memory consumption that each patch set incurs, I’ve added two additional benchmarks to our normal set of benchmarks.

No BDB

This is a custom build of EDS gnome-3-6 branch with the removal of the BDB usage in the local file backend.

At this point there is no extra table in the SQLite to handle multi-valued vCard attributes, it’s simply a comparison of storing the vCard data in the BDB vs SQLite only.

Custom Light

This is a special run of our regular openismus-work branch, but with only the “Full Name” configured and indexed in the summary.

So this benchmark is a light-weight summary with considerably less columns (and one less table) used in the SQLite.

I ran this variation in the suspicion that SQLite might require significantly more memory with the additional multi-value table created to handle multi-valued attributes such as E_CONTACT_TEL.

Benchmark Results

Note that the RSS and VMS memory snapshots are taken by way of observing the /proc/$pid/status file for both the EDS server process and client benchmark process directly after stopping the clock for each benchmark in the suite. So a given value in the charts presented below is based on the “VmRSS” value of the server process added to the “VmRSS” value of the client process.

First, let’s show the results, or at least some of them, to put our deductions into context:

50 Contacts

50 Contacts

… Skipping a few results here in the interest of avoiding clutter … lets jump directly to 400 contacts …

memory-usage-rss-00400

400 Contacts

800 Contacts

800 Contacts

1600 Contacts

1600 Contacts

3200 Contacts

3200 Contacts

6400 Contacts

6400 Contacts

12800 Contacts

12800 Contacts

And now, some of the conclusions I came to while observing the results

BDB Removal

When compared to the unmodified EDS 3.6 branch, we can observe that the BDB removal reduces memory consumption for most reasonably sized address books. Up until we run the benchmark for 3,200 contacts, memory consumption is less without BDB… with 3,200 contacts and higher, memory consumption is increased by removing the BDB.

Without an in depth understanding of SQLite internals, I think we can deduce that the SQLite starts to require more memory to handle databases with >= 3,200 rows

Custom Light

This benchmark basically disproved my suspicion.

While using exactly the same code-base as the “EDS Custom” and “EDS Custom DRA” benchmarks; Using more indexes and tables in the SQLite does not seem to incur much of a difference in terms of memory consumption.

While the output is certainly different, as specially with large addressbooks, I don’t see much of a noticeable pattern here.

EDS Custom

This benchmark is basically the openismus-work branch with fully customized indexes for better performance in telephone number lookups.

When comparing this one to the unmodified EDS 3.6 benchmark, we can observe that memory consumption is slightly less using the custom EDS code than stock EDS 3.6.

When comparing this to the removal of BDB, we can notice that, as specially for small addressbooks, the base memory requirement of the EDS Custom is significantly higher than with only the BDB removal.

This second point is easily explainable, since removal of BDB alone reduces the overall memory footprint of EDS. The custom EDS benchmarks, without actually leveraging the Direct Read Access mode still links against the EDataBook library. Essentially this replaces the memory footprint overhead incurred by linking to BDB with a different overhead incurred by linking directly with EDataBook.

EDS Custom DRA

This benchmark is particularly interesting.

For smaller addressbooks the Direct Read Access mode indeed costs more resident memory than any other benchmark. This can be attributed at least partly to the penalty of loading an EDataBook into memory on the client side. Consequently, loading the EDataBook also loads the backend module in the client process, meaning we also have a running EBookBackendFile in the client as well as client side linkage and usage of the SQLite library.

However, once we approach addressbooks with 1600 contacts and more, the overall resident memory consumption starts to even out. Direct Read Access mode actually costs significantly less than any other benchmark for addressbooks as large as 6400 contacts and more.

These results are a bit harder to explain. My theory is that since the EDS server process essentially goes to sleep after adding the initial contacts. All queries thereafter require no interaction with the EDS server process.

Some things to consider here are:

  • ┬áThe cost in memory of constantly waking up the EDS process to handle a query
  • The cost of server side heap allocations used to deliver the results over D-Bus
  • The cost of client side heap allocations used to receive results over D-Bus

Overall Memory Consumption differences

In summary, we can conclude that after all measures taken to improve performance of contact fetches in EDS; the Direct Read Access mode is the single element which makes a tradeoff in terms of memory consumption versus speed.

Without the Direct Read Access patches, memory consumption as well as time to fetch contacts has seen a net improvement. With Direct Read Access enabled we see that for smaller address books an additional memory overhead is required, while with larger addressbooks (larger than 3,200 contacts); overall resident memory usage has seen a significant improvement as well.