FastMM4 large memory allocation–benchmarking VirtualAlloc

[Primož Gabrijelčič 223]

 

Earlier this week a long-time customer asked me why FastMM allocates large memory blocks from the top of the memory and if that option could safely be turned off. Their reasoning was that such allocations are much slower than normal ones. The question surprised me as I didn’t know of any such difference in speed so I did what I usually do–I wrote a benchmark application and measured it.

TL;DR Yes, allocating from the top is slower. No, the difference is not big and in most cases you’ll not even notice it.

There were, however, other interesting results that my simple benchmark pointed out. More on that in a moment, but first…

Allocating from bottom and top

In Windows, the code can ask for a memory block by calling VirtualAlloc with flag MEM_COMMIT and Windows will give you a suitable chunk of memory. This memory will usually be allocated from the start of the virtual memory visible to the program.

The code can, however, call VirtualAlloc with flag MEM_COMMIT OR MEM_TOP_DOWN and Windows will return a block from the end of virtual memory available to the process. In a typical 32-bit Delphi program first such memory block will have address close to $7FF00000 (but a bit lower).

You may want to allocate memory “from the top” if your program has two very distinct modes of allocating memory and you don’t want to mix them. For example, a frequently reallocated memory could come “from the bottom” and large blocks that are used for long periods of time “from the top”. This can reduce memory fragmentation, but the potential advantages are specific to each program. In other words – maybe it will help, maybe it will hurt.

Another good scenario for MEM_TOP_BOTTOM is testing 64-bit code ported from 32-bits. For example, a typical “from the top” allocated block in a 64-bit program will have an address like this: $7FF4FDE30000. If your code at some point stores pointers into 4-byte integers, part of the address will be lost and as soon as that integer is converted back into a pointer and the code accesses that pointer, you’ll quite probably get an access violation. If a memory comes “from the bottom”, such problems would not be so easily detected.

FastMM4 allocates large blocks (with sizes greater or equal to 258.5 KB) “from the top”. If I recall correctly, this was done to prevent memory fragmentation. Additionally, it can allocate all other block “from the top” if you define conditional symbol AlwaysAllocateTopDown and rebuild. (You have to use FastMM4 from github instead of the built-in Delphi version to use this functionality.) You can use this mode to test 32-bit programs ported to 64-bit code.

MEM_TOP_DOWN is slower?

The article my customer pointed to claimed that allocating from the top works much slower than allocating from the bottom. Even worse, the allocation algorithm was supposed to work in O(n^2) time so each additional allocation needs more time to execute. To top that off, the official documentation for the MEM_TOP_DOWN flag mentions:

This can be slower than regular allocations, especially when there are many allocations.

To verify that claim, I wrote a trivial benchmarking app (download it here). It allocates from 1 to 6000 blocks of size 264752 and measures the time needed. Block size 264752 was picked because at that size FastMM4 starts allocating memory “from the top”. 6000 blocks can safely be allocated in a 32-bit application (6000 * 264752 = 1.5 GB). In my tests I could allocate 6105 such blocks before I ran “out of memory” but just to be on the safe side I reduced the number in the released application.

Results, measured on my fresh new notebook with a i7-8750 processor, were much closer to my expectations than to some O(n^2) algorithm. The “Top” algorithm is slightly slower (needs more time to execute) but the difference is not drastically large.
 

What’s going on then? Is MEM_TOP_DOWN slow or not?

As it turned out, the article I was referring to was written in 2011 and Windows have improved a lot since then. I don’t know which Windows version has fixed the “top allocation” problem, but it definitely doesn’t appear in Windows 7 and 10.

Another interesting result is that the first 200 MB (approximately) are almost “free”. Somewhere around that number, the execution time jumps from around 3 ms to 50 ms and then continues to grow in more-or-less linear fashion. The benchmarking program measures each test only once and is therefore very susceptible to measurement errors but the result clearly shows an O(n) algorithm.

Why are allocations smaller than 200 MB so fast? I’m guessing that Windows maps such amount of physical memory into the process’ virtual space when the process is started. When you exceed that limit, the allocator needs more time to allocate physical memory and map it into the process’ virtual space. That’s, however, just a guess. If you know better, please let me know in the comments.

How fast are YOUR allocations?

Just for the sake of completeness I rerun tests on my main PC (HP z820 with two E5 Xeons) and the results completely surprised me.
 

The shape of the curve is almost the same–but notice the difference in speed! On the laptop, 4000 allocations execute in 250 ms. On the Xeon machine, over 1000 ms is needed for the same job.

This machine is quite old (around 4 years IIRC) and it obviously contains a much slower memory. I know that computers can have faster or slower memory chips, but I never expected to see such a big difference in VirtualAllocspeed. (And yes, both machines are running latest Windows 10.)

Now the whole shebang started to interest me even more, and I did some further tests on a few PCs used by fellow programmers. All of them were running Windows 10. As you can see below, there is some difference between them but none are so slow than my main computer 😞 Maybe the time has come to upgrade…
 

If you want to download raw data and compare it to your own results, you can access it here.

MEM_TOP_DOWN or not?

The difference in speed is not that big–and most programs will not notice it–but I have to agree with the customer. The time has come to remove hard-coded MEM_TOP_DOWN from FastMM4 and replace it with a conditional {$IFDEF AllocateLargeBlocksTopDown}MEM_TOP_DOWN{$ENDIF}.

 

I have created pull request for that change: https://github.com/pleriche/FastMM4/pull/75

 

(Original blog post: https://www.thedelphigeek.com/2019/04/fastmm4-large-memory.html)

Edited April 27, 2019 by Primož Gabrijelčič
Added link to original post

[Ian Branch 128]

Slightly of-topic but related.  Is there any benefit to using the full FastMM4 as opposed to just the in-built Delphi version?

[Hallvard Vassbotn 3]

Yes, you get access to many tweaks and debug options, including full stack trace of memory leaks. 

 

From 2007:

http://hallvards.blogspot.com/2007/05/use-full-fastmm-consider-donating.html?m=1

[Primož Gabrijelčič 223]

... or my recent additions for finding allocation/deallocation bottlenecks: https://www.thedelphigeek.com/2016/02/finding-memory-allocation-bottlenecks.html

 

You'll also get much better leak reporting. That alone is worth using the fresh github version.

[AlekXL 8]

I am wondering, in that possible to use FullDebugMode features of FMM4, in Lazarus/FPC? At least on Windows?

[Primož Gabrijelčič 223]

An interesting result from Windows 2012 Server, dual E5-2620v3 @ 2.4 GHz.

 

 

MEM_TOP_DOWN is indeed much slower here - but only compared to the same machine. All allocations are blindingly fast compared to Windows 10 computers I've tested before. Windows Server is quite obviously optimized for different usage than Windows 10.

[Микола Петрівський 10]

In my tests 64-bit app is ten times faster, then 32-bit on real hardware. In VM difference is only two times, but 32-bit code is twice faster, then on host. Very interesting results.