slab cache

I am trying to improve the speed of my fs code. I have a fixed sized buffer for my fs, I currently use kmalloc for allocation of buffers greater than 4k, use get_free_page for 4k buffers and vmalloc for large buffers.

Hi,

Quoted 2 lines · David Chow

I am trying to improve the speed of my fs code. I have a fixed sized buffer for my fs, I currently use kmalloc for allocation of buffers greater than 4k, use get_free_page for 4k buffers and vmalloc for large buffers.

Allocations larger than pagesize always put a higher stress on the VM and reduce performance. Your best bet for top performance will be simply to perform no allocations larger than pagesize. You can use a slab cache for those allocations if you want, and that may have some advantages depending on the locality of allocations in your code.

Using 4k buffers does not limit your ability to use larger data structures --- you can still chain 4k buffers together by creating an array of struct page* pointers via which you can access the data.

--Stephen

Quoted 3 lines

Using 4k buffers does not limit your ability to use larger data structures --- you can still chain 4k buffers together by creating an array of struct page* pointers via which you can access the data.

Yes, but for me it is very hard. When doing compression code, most of the stuff is not even byte aligned, most of them might be bitwise operated, it need very change to existing code.

get_free_page to allocate memory that is 4k to avoid some stress to the vm, I have no idea about the difference of get_fee_page and the slab cache. All my linear buffers stuff is already using array of page pointers, if there any benefits for changing them to use slabcache? Please advice, thanks.

Hi,

Quoted 7 lines · David Chow

Quoted 4 lines

Using 4k buffers does not limit your ability to use larger data structures --- you can still chain 4k buffers together by creating an array of struct page* pointers via which you can access the data.

Yes, but for me it is very hard. When doing compression code, most of the stuff is not even byte aligned, most of them might be bitwise operated, it need very change to existing code.

Perhaps, but the VM basically doesn't give you any primitives that you can use for arbitrarily large chunks of linear data; things like vmalloc are limited in the amount of data they can use, total, and it is _slow_ to set up and tear down vmalloc mappings.

Quoted 2 lines

get_free_page to allocate memory that is 4k to avoid some stress to the vm, I have no idea about the difference of get_fee_page and the slab cache. All my linear buffers stuff is already using array of page pointers, if there any benefits for changing them to use slabcache? Please advice, thanks.

It might be if you are allocating and deallocating large numbers of them in bunches, since the slab cache can then keep a few pages cached for immediate reuse rather than going to the global page allocator for every single page. The per-cpu slab stuff would also help to keep the pages concerned hot in the cache of the local cpu, and that is likely to be a big performance improvement in some cases.

--Stephen

Thanks for comment, since you mention about cache, do you mean CPU L2 caches? I don't use to dynamic alloc and dealloc pages, I have a fixed sized cache per CPU, even using vmalloc I will only do it only once during module initialize, and dealloc only on unload,