NUMA aware heap memory manager

NUMA aware heap memory manager

How to use our resources wisely in multi-thread and multi-proccessor systems

Michael ShteinbokShai Ben Nun

Supervisor: Dmitri Perelman

What’s the problem?While increasing number of cores/processors

seems to increase the performance in proportion, it usually increases in lower numbers and sometimes even slows us down.

Ok, But Why?Some physical issues can still cause a

bottleneck. For example Memory Access and Management.

UMA – Uniformed Memory Access

Solution - NUMA Non Uniform Memory Access - each CPU has it’s

own close memory with quick access. One processor can access the other memory units

but with slower BW. The NUMA API lets the programmer to manage the

memory.

How One Manages The Memory?When we allocate and free, we are blind to

the process in the background. For this need we have

different lib function. (glibc, TCMalloc, TCMalloc NUMA Aware)

The article refers to this management problem and offers the TCMalloc NUMA Aware heap manager

http://developer.amd.com/Assets/NUMA_aware_heap_memory_manager_article_final.pdf

Performance Change The Previous solutions(more BW, more local

access):

Project Goals Improving the current TCMalloc numa aware

in scenarios it losses performance Managing wisely the memory of different

cores and offer a new “read” function that will be faster (TBD)

Problem In Current SolutionWhen thread frees memory that was allocated by another

thread (on another numa node) we can get performance loss

X = malloc

Thread A

Thread AFree pool

Free (X)

X = Allocate

Thread A

Free (X)

Thread B

Our BenchmarkEach couple of threads on a different numa

node. 8 quad-cores processors => Total 16 couples.

Allocator Thread

Free Thread

Alloc Rand List (5000)

Access Memory

AllocQueue

First-touch policy

Achieving Project Goals Step By Step

Learn how does the TCMalloc workFind scenarios that makes current TCM to

loss performance.Sort the scenarios by most likely to occur in

real environmentImplement a support in these scenarios

Make new TCMalloc to get better performance!