Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ur] optimize align allocation in DisjointPool #625

Merged
merged 2 commits into from
Jun 22, 2023
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 28 additions & 3 deletions source/common/uma_pools/disjoint_pool.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,13 @@ static void *AlignPtrUp(void *Ptr, const size_t Alignment) {
return static_cast<char *>(AlignedPtr) + Alignment;
}

// Aligns the value up to the specified alignment
// (e.g. returns 16 for Size = 13, Alignment = 8)
static size_t AlignUp(size_t Val, size_t Alignment) {
assert(Alignment > 0);
return (Val + Alignment - 1) & (~(Alignment - 1));
}

DisjointPoolConfig::DisjointPoolConfig()
: limits(std::make_shared<DisjointPoolConfig::SharedLimits>()) {}

Expand Down Expand Up @@ -272,6 +279,9 @@ class DisjointPool::AllocImpl {
// Configuration for this instance
DisjointPoolConfig params;

// Coarse-grain allocation min alignment
size_t ProviderMinPageSize;

public:
AllocImpl(uma_memory_provider_handle_t hProvider, DisjointPoolConfig params)
: MemHandle{hProvider}, params(params) {
Expand All @@ -285,6 +295,12 @@ class DisjointPool::AllocImpl {
Buckets.push_back(std::make_unique<Bucket>(Size2, *this));
}
Buckets.push_back(std::make_unique<Bucket>(CutOff, *this));

auto ret = umaMemoryProviderGetMinPageSize(hProvider, nullptr,
&ProviderMinPageSize);
if (ret != UMA_RESULT_SUCCESS) {
ProviderMinPageSize = 0;
}
}

void *allocate(size_t Size, size_t Alignment, bool &FromPool);
Expand Down Expand Up @@ -742,9 +758,18 @@ void *DisjointPool::AllocImpl::allocate(size_t Size, size_t Alignment,
return allocate(Size, FromPool);
}

// TODO: we potentially waste some space here, calulate it based on minBucketSize and Slab alignemnt
// (using umaMemoryProviderGetMinPageSize)?
size_t AlignedSize = (Size > 1) ? (Size + Alignment - 1) : Alignment;
size_t AlignedSize;
if (Alignment <= ProviderMinPageSize) {
// This allocation will be served from a Bucket which size is multiple
// of Alignment and Slab address is aligned to ProviderMinPageSize
// so the address will be properly aligned.
AlignedSize = (Size > 1) ? AlignUp(Size, Alignment) : Alignment;
} else {
// Slabs are only aligned to ProviderMinPageSize, we need to compensate
// for that in case the allocation is within pooling limit.
// TODO: consider creating properly-aligned Slabs on demand
AlignedSize = Size + Alignment - 1;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not exactly sure what's the purpose of adding alignment to the size here? Can't we just always AlignUp? This should always be called with size being a multiple of alignment anyway (which is what AlignUp enforces).

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider a case where size = alignment = 1MB. Slabs are only aligned to ProviderMinPageSize (let's say 4KB) so aligning up to Alignment is not enough (and we're allocating from Slabs up to 2GB allocations).

When we're allocating directly from the memory provider (line 775) we don't need to do any adjustments, we just leave it to the provider.

Copy link
Contributor

@pbalcer pbalcer Jun 21, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is potentially very wasteful if there are many such allocations. We just need to create a bucket that offsets the beginning of the slab. Or we can just create an aligned slab, which should be doable if alignment is page-aligned. And even if it's not, at worst we'd waste 4kb per slab.
Let's leave it as-is for now, but might be worth investigating whether this is a common case and worth optimizing.

But please add a comment here explaining this.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done. Just one clarification - the 2GB I mentioned is actually just the upper limit for poolableSize. By default poolableSize is around a few MB (so allocation bigger than a few MBs will bypass the pooling).

}

// Check if requested allocation size is within pooling limit.
// If not, just request aligned pointer from the system.
Expand Down