Memory – Part 6: Optimizing the FIFO and Stack allocators

Introduction

The most used custom allocators at Intersec are the FIFO and the Stack allocators, detailed in a previous article. The stack allocator is extremely convenient, thanks to the t_scope macro, and the FIFO is well fitted to some of our use cases, such as inter-process communication. It is thus important for these allocators to be optimized extensively.

We are two interns at Intersec, and our objective for this 6 week internship was to optimize these allocators as far as possible. Optimizing an allocator can have several meanings: it can be in terms of memory overhead, resistance to contention, performance… As the FIFO allocator is designed to work in single threaded environments, and the t_stack is thread local, we will only cover performance and memory overhead.

Tools used

To be able to quickly assess the impact of our changes, we began by writing a quick benchmark that uses the allocator to make a few million allocations and measures the time it takes.

For the FIFO allocator, the scenario is the following: in the first loop, we request memory for the allocator and we immediately release it. In the second loop, we loop several times through an array of pointers, and at each iteration, the pointer is freed (if it is not NULL) and then reallocated. The effect of this is that there is a constant amount of memory allocated at all times, and we keep a FIFO behaviour: the first pointer allocated is also the first freed. The third loop still uses the array of pointers, but instead of looping through it regularly, it selects a random one at each iteration, frees it and re-allocates it. This way we still have a constant amount of memory allocated, but we lose the FIFO behaviour.

For the stack allocator the scenario is simpler: a recursive function allocates some memory, calls itself up to a given recursion depth, allocates some more memory after the recursive call ended and then returns. The t_scope mechanism ensures that all the memory allocated by a function is freed when it returns.

To look for performance bottlenecks, we extensively used the linux perf tool which allows to quickly see how our changes affect crucial parameters such as the number of branch misses or cache misses, and where these events originate from. Since the allocation routines are almost flat functions, we needed an insight down to the instruction level.

We also instrumented both allocators to keep track of several statistics such as the memory allocated, the memory used, the numbers of memory requests, the number of subsequent calls to mmap() or calloc(), and the time the different functions (malloc(), free(), realloc()) take to complete. This instrumentation also allowed us to check the behaviour of our allocators in real products, and not only in our benchmark. This instrumentation incrementally gathers all the necessary usage information, and periodically dumps it into a file.

Using this tool, we have analysed the main memory allocation patterns in Igloo, the Intersec Geo-Locator. The demonstration program for Igloo has been used to get a simplified real-world use-case. For both allocators, an entry in the file dump is issued each 300 simple operations (malloc(), free(), realloc()), and at each long operation (memory request from the system, memory page cleaning…). Feeding only 10,000 events per second for 5 minutes into Igloo already produces an overwhelming quantity of dumps, around 6GB. Exploiting this quantity of dumps has been a bit of a challenge, but has provided some insight into the effective usage of the allocators: for instance, we get around 100,000 allocations per minute.

The FIFO allocator

The FIFO allocator is relying on large pages of memory allocated through mmap(). The pages are split into variably-sized blocks as requested. Each block maintains a reference to the page it belongs to (actually the offset of the block in the page), as well as its size. The FIFO pool keeps a reference to the current page as well as a free page, which is not always set but is used to reduce the number of page deallocation / reallocation operations.

FIFO allocator

Representation of the FIFO allocator

With this structure, the behaviour of the 3 main functions (alloc(), realloc(), free()) is straightforward:

  • For alloc(): if there is enough space in the current page, return a pointer to the beginning of the free space in the current page (fast path), else allocate a new page.
  • For realloc(): if the new size is smaller than the original size, return the same pointer. Else if the pointer we are reallocating is the last block of its frame and there is enough space in the frame, return the same pointer too, else allocate a new block, copy the memory from the old block to the new and free the old one.
  • For free(): decrease the number of blocks in the containing page. If this number is zero, destroy the page.

The FIFO allocator always returns memory initialized to zero. It also aligns all memory blocks on multiples of at least 8 bytes.

For these three functions, the fast code path (without operations on pages) is extremely fast, and the slow path is not supposed to happen too often (as long as the default page size is big enough to hold a reasonable number of allocations, which is the responsibility of the owner of the pool) but as we will see, there was still room for improvement. Trying to reduce the frequency of the slow paths and increasing the speed of the fast path allowed us to make the FIFO allocator 30% faster than it originally was.

Page management

The most significant improvements we have made to the FIFO allocator were achieved by improving the page management policy. What was initially done was:

  • The pool is created without a page (current = NULL)
  • When allocating or reallocating, if there is not enough room in the current page, request a new page.
  • When freeing, if we are freeing the last block of a page:
    • If the page is the current one, set current to NULL
    • If there is already a freepage, delete our page, else reset the page and save it as freepage.

Requests for a new page are performed through mem_page_new(), which returns the freepage if it exists, and allocates a new one otherwise. An important thing to note is that the free() function is responsible for the deletion of unused pages: the pool doesn’t keep references to any page unless it is the current page. Instead, each page tracks the number of blocks still alive in it, and when all blocks are freed, this count reaches zero and the page must be freed.

FIFO allocator - old behaviour

Initial code flow of alloc() and free() – Click image to enlarge

This scheme is good, but it has several drawbacks: for instance, let’s consider the case where there is only one pointer allocated at a time. When it is freed, we set current to  NULL, and then we reset the page (assuming that there is no freepage, which is the case after the first iteration), i.e. we memset() the used space of the page to zero. As a consequence, each time we want to allocate, we need to fetch the free page first. This means that in this case, all calls trigger slow path runs.

To circumvent it, we gave the current page a special treatment in free(): if its last block is freed, we do nothing, so that the next allocations will keep filling it. However, we also had to add an extra check in the slow path of alloc(): when the current page is out of space, if it turns out that there is no alive block in it, we have to reset it (and then we can reuse it).

This has the side effect that current is never NULL once the first page has been created. If we allocate the first page at the creation of the pool, we can be sure that current will never be NULL, thus we can remove this check from the fast path of both alloc() and realloc().

Another improvement that was made, which is closer to a trade-off between memory overhead and performance, is the following: if a page becomes available, and we already have a freepage, instead of deleting it straight away, we check the remaining space available in the current page: if the free space in the current page is less than 1/8th of the size of the available page, we immediately replace it with the available page. This doesn’t incur a raw performance gain in the benchmark, but it tackles a bad behaviour that we spotted in a product where a page would be deleted just before a new one was allocated. This change greatly reduces the page allocation frequency in this case.

FIFO allocator - new behaviour

New code flow of alloc() and free() – Note: mem_page_new() has not changed – Click image to enlarge

Page allocation

The allocation of new pages was initially done through calls to mmap(), which always returns memory initialized to zero. However, using calloc() turned out to be significantly faster. Here are the results of the time it took for our benchmark to complete, depending on page size, for mmap() and calloc():

Comparison of calloc() and mmap() for page allocation

Comparison of calloc() and mmap() for page allocation

Both functions are approximately as fast for 1MiB pages, which (according to the malloc() source) is the threshold above which malloc() switches to mmap(). But surprisingly, for bigger pages calloc() is again faster than mmap(), so we switched to calloc() for all page sizes. The default page size being 64KiB, this simple switch made our benchmark run faster by 10%.

Other optimisations

A conventional optimisation technique that we went through is structure packing. Because of how the memory pools are managed, each pool structure has to inline a 40-byte mem_pool_t structure, which leaves 24 bytes on the first cache line for the specific pool data. The current and freepage pointers consume 16 of these 24 bytes.

The FIFO pool has a specific way of being destroyed: if there is still allocated memory in the pool when we try to delete it, it stays alive until all the memory is freed. Because of this, it has to keep a boolean indicating if the destructor has been called (and forbidding allocations if it is the case) and the number of pages allocated. It also keeps the default page size which is configured at the creation of the pool. Moreover, it keeps track of the total memory allocated, and of the currently used memory, which are needed by statistics reporting functions. This makes a lot of information for the 8 remaining bytes, however among these members, the only ones that are used in the fast paths are the alive boolean and the occupied integer. occupied can’t be stored on 4 bytes, because it may very well exceed 4GiB (it often does so in the benchmark), but if we use an 8-byte integer, then there is no room left for alive on the first cache line. The fastest solution here was to use bitfields: 63 bits for occupied and 1 for alive. This gave a ~3% speed improvement in the bench.

The resulting structure is the following:

Another optimisation we did is actually quite funny: consider the two code blocks below, taken from alloc():

Can you tell which one is the fastest? No you can’t, you would need to compile the whole code to do so. The thing is that thanks to the perf tool, we saw that the assembly instruction corresponding to page = mfp->current; was taking a lot of time. We tried moving it around, and we saw that putting it before the size assignment yielded a 2% improvement in the bench. Upon closer inspection of the assembly code generated, we saw that this modification completely changed GCC’s register allocation, and thus reduced register interlock in this portion of the code.

The Stack allocator

The Stack allocator, unlike the FIFO allocator, is frame-based: allocations are grouped in frames, to allow batch deallocation. Frames are stacked in a LIFO ordering. Prior to a group of allocations, one must push a frame into the allocator, and pop it to free the group. The frame object records all the bookkeeping information for allocations. It still shares some properties with the FIFO allocator, being block-based (a block here denotes what the FIFO allocator called a page). This way, it can achieve very fast size-agnostic allocations. The model is simple: the frame holds the current position in the block. Upon a memory request, we just need to bump this pointer forward. Thanks to the frame-based model, deallocation is almost free, we just need to pop the last frame. The allocator will recover its exact state, saved by the matching push. The third meaningful operation is realloc(). General purpose reallocation of an arbitrary memory chunk is in general difficult and slow. However, the block-based model allows a very useful case: reallocating the last allocation is almost free. We just need to update the position pointer.

Stack allocator

Representation of the stack allocator

Since the fast path is already very efficient (16 cycles), the actual improvement resides mostly in controlling the frequency of slow paths, i.e. allocation of new blocks from system.

Block allocation policy

The block allocation policy is constrained by the need for very fast deallocations. This incurs that almost all the block handling is done on demand, at allocation time. The general block management uses a simple list of blocks. At all times, all the blocks before the current frame’s are in use, and all the latter blocks are free. When a block is required, the list is traversed. If no suitable block is found a new one is allocated. Because of the typical size of the blocks (at least 64KiB), the creation of new blocks is made via a call to the libc’s malloc(), which in turn yields a call to mmap(). This incurs a substantial latency, so it is important to reuse the blocks. This reduces memory consumption, memory fragmentation and allocation time.

Compared to the fast path of the allocator, searching a new block is very time-consuming, and creating one is all the more. A straightforward manner to reduce the penalty is to be eager with resizes. For instance, take the following diagram: the current frame (Frame 2) owns an almost full block, so pushing a new frame will require a new block. When pushing Frame 3, the code used to consider that it must not alter the former, so Frame 2 still points to Block A, and Frame 3 to Block B. After several operations, this frame will be popped, and this last block will be put back on free list, back to square one. Then, another frame will be pushed, and require that same free block. This forgetful behaviour is not the most efficient.

A first step towards solving this is to always update the current frame when a block is added. This way, on the example below, Frame 2 remembers it will soon need Block B, and avoids repeated block fetching. This furthermore allows to save a branch in the allocation code. A second step is to add speculative block replacement: whenever a block is added to the top frame, it can conditionally be added to the previous frames too.

Example situation

Example situation before pushing a new frame. The current frame (Frame 2) points to the almost full block A.

Block searching

When a block was needed (slow path), the block search was greedy:

  • visit the free blocks list, if the block is big enough return it, else destroy it
  • if no suitable block is found, create a new one

Actually, this scheme just considered that the small blocks will always be too small for further allocations. While this keeps memory usage bounded and aims at reducing the fragmentation, it relies on the assumption of fairly regular request sizes. This behaviour is deadly when an oversized allocation slips in. For instance, this pattern of allocation:

Using our instrumentation of the allocator, we created the following graph, plotting the current memory allocated from the t_stack (black) and the memory available in blocks (purple). Horizontal axis is the number of file dumps, vertical axis is the size in bytes.

Original behaviour for worst_case()

Original behaviour of the allocator for worst_case(), initial depth 20, truncated at 2500 allocator operations. Cumulated size of block allocations: 35MiB

Analysing this graph, we observe that the allocator allocates a bunch of small blocks for the first scope, puts them on the free list when popping the frame, wipes everything to create a 80kB+ block, reallocates a similar bunch of small blocks, and repeat this all over. This is clearly suboptimal.

In order to avoid this, we chose to stop the block destruction as soon as we reached the requested size. This keeps the bounded memory usage and still reduces memory fragmentation. On this very example, the gain in time is around 30%.

Corrected behaviour for worst_case()

Corrected behaviour of the allocator for worst_case(), initial depth = 20, truncated at 2500 allocator operations. Cumulated size of block allocations: 7.5MiB

Block sizing

In order to bound the block switch frequency, those are sized proportionally to average of the size of the requests. A newly allocated block must be at least 64 times the mean allocation size as computed below.

This method for computation is great for use in the allocator. It uses basic integer operations only (addition, comparison and shift), and be updated in a few clock cycles, and most of all is smooth. The resulting average converges roughly at the same speed than an exponential moving average with decay 1/27279, but is 25% smoother (computed from the expected deviation after an new input value).

Still, some parts of this algorithm could be elided by some maths. First, the maximum allocation allowed is 1GiB. This check is primarily made for sanity reasons, to avoid devious allocations caused by unchecked overflow. Since av->nb is smaller than UINT16_MAX, av->sum will always be smaller than 1GiB * UINT16_MAX < UINT64_MAX. Hence, the overflow check is always false.

Removing this check does not improve anything, it just removes a jump instruction that was always predicted. The other test we may want to remove is the truncation to 128MiB. A probability analysis shows that the expected oversize for a single peak p is roughly equal to p / 550, for any average request size. After an allocation of 1GiB, the expected overhead in block size will be 2MB. This is fairly tractable after a 1GiB allocation, so the truncation can be removed.

Average filters

Average filters on a 4M allocation sequence, with mean 16KiB, Gaussian noise of standard deviation 3KiB, 5 bunches of 2000 allocations of 256KiB, and a single peak allocation of 1GiB

Fast path improvement

Besides being already very efficient, the preferred path for allocation contained some overhead in the way alignment was computed.

Memory alignment is a central performance problematic. Most processors access memory with granularity of a word, which means that any read or store will at least fetch this size. From the way computer memory is handled, these accesses work best when addresses are word-aligned. Moreover, some recent instructions (and some old architectures) do require aligned operands (vector operations in SSE2 require 128-bit alignment).

The easiest way to fulfil alignment requirements in an allocator is to keep uniform alignment for all the allocated objects. This takes the form of rounding up the size to next boundary, an operation happily optimized by the compiler, since the alignment is known at compile time. Although being most efficient, the memory overhead of this scheme can be unbearable: for 16-byte alignment, allocating a single integer takes four times the size!

This behaviour has fortunately been changed in Intersec allocators in favour of on-demand alignment. This model requires to round up the returned pointer to demanded boundaries. In the original code, both the returned address and the size were rounded up. This accounts for unnecessary checking:

  • for sizes computed with sizeof(), the alignment is guaranteed by the C language (C99 standard, paragraph 6.5.3.4/3)
  • the processor only cares about address alignment

The key for efficient computation is to use the fact that alignment is always a power of two. The choice in implementation resides in the storage of the alignment boundary: either in full form (like the output of alignof() operator, a power of two), or in logarithmic form (the position of the bit).

The full form has two advantages: first, it requires a smaller and simpler assembly, faster to execute, and it does not need to compute the logarithm of the alignment. Eventually, benchmarking shows a 9% improvement in the allocation code.

realloc()

In this kind of block-based allocator, the realloc() operation for last allocated block is very cheap. Indeed, at each allocation, we record the returned address. At realloc() time, it is just necessary to bump the position pointer a little farther. Since the work on the block allocation policy already impacts the realloc() function, the only part where improvements can be made is the realloc() code itself.

We followed a logic of branch-elimination on the function. The most used path is in red on the graph below. The typical use of this function is the growing of dynamic arrays: it is expected to be most efficient. This refactoring allows for better branch optimization, and results in a 10% gain in reallocation speed.

Old vs new realloc() - Click image to enlarge

Old vs New realloc() code path – Click image to enlarge

Overall gains

In order to evaluate the consequences of our work on the stack allocator, we considered benchmarking both the original and our final versions of the code. Measuring raw allocation speed for different allocation / deallocation ratios show a consistent 17% gain. Still, this benchmark only considers a very regular allocation pattern, and does not account for exceptional use cases like the worst_case() function above. Thanks to this observation, this former gain is more of a lower bound on the improvements.

Conclusion

These allocators were already quite efficient, but we still managed to find ways to improve them, the most significant of which are detailed in this article. By grabbing a few percent here and there, we eventually improved the overall performance of the two allocators by around 20-30%, depending on the use cases. It was a challenging task, because of the complexity of the matter as well as the fact that we had to dive into a particularly developed code base.

We learnt a lot of things during these six weeks, and we are particularly happy to have done so at Intersec, where skill meets cheerfulness.

Camille Gillot & Aloÿs Augustin, interns @ Intersec
Thanks to Jonathan Squirawski, our tutor, for his patience and careful review.
Thanks to Florent Bruneau, for his advice and his review.