Hi all, We recently worked on GTPU tunnel and our target is to create 2M tunnels. It is not as easy as it looks like, and it took us quite some time to figure it out. The biggest problem we found is about mheap, which as you know is the low layer memory management function of vPP. We believe it makes sense to share what we found and what we've done to improve the performance of mheap.
First of all, mheap is fast. It has well-designed small object cache and multi-level free lists, to speed up the get/put. However, as discussed in the mail list before, it has a performance issue when dealing with align/align_offset allocation. We managed to locate the problem is brought by a pointer 'rewrite' in gtp_tunnel_t. This rewrite is a vector and required to be aligned to 64B cache line, therefore with 4 bytes align offset. We realized that it is because that the free list must be very long, meaning so many mheap_elts, but unfortunately it doesn't have an element which fits to all 3 prerequisites: size, align, and align offset. In this case, each allocation has to traverse all elements till it reaches the end of element. As a result, you might observe each allocation is greater than 100000 clocks/call with 'show memory verbose'. It indicates the allocation takes too long, while it should be 200~300 clocks/call in general. Also you should have noticed 'per-attempt' is quite high, even more than 100. The fix is straight and simple : as discussed int his mail list before, to allocate 'rewrite' from a pool, instead of from mheap. Frankly speaking, it looks like a workaround not a real fix, so we spent some time fix the problem thoroughly. The idea is to add a few more bytes to the original required block size so that mheap will always lookup in a bigger free list, then most likely a suitable block can be easily located. Well, now the problem becomes how big is this extra size? It should be at least align+align_offset, not hard to understand. But after careful analysis we think it is better to be like this, see code below: Mheap.c:545 word modifier = (align > MHEAP_USER_DATA_WORD_BYTES ? align + align_offset + sizeof(mheap_elt_t) : 0); bin = user_data_size_to_bin_index (n_user_bytes + modifier); The reason of extra sizeof(mheap_elt_t) is to avoid lo_free_size is too small to hold a complete free element. You will understand it if you really know how mheap_get_search_free_bin is working. I am not going to go through the detail of it. In short, every lookup in free list will always locate a suitable element, in other words, the hit rate of free list will be almost 100%, and the 'per-attempt' will be always around 1. The test result looks very promising, please see below after adding 2M gtpu tunnels and 2M routing entries: Thread 0 vpp_main 13689507 objects, 3048367k of 3505932k used, 243663k free, 243656k reclaimed, 106951k overhead, 4194300k capacity alloc. from small object cache: 47325868 hits 65271210 attempts (72.51%) replacements 8266122 alloc. from free-list: 21879233 attempts, 21877898 hits (99.99%), 21882794 considered (per-attempt 1.00) alloc. low splits: 13355414, high splits: 512984, combined: 281968 alloc. from vector-expand: 81907 allocs: 69285673 276.00 clocks/call frees: 55596166 173.09 clocks/call Free list: bin 3: 20(82220170 48) total 1 bin 273: 28340k(80569efc 60) total 1 bin 276: 215323k(8c88df6c 44) total 1 Total count in free bin: 3 You can see, as pointed out before, the hit rate is very high, >99.9%, and per-attempt is ~1. Furthermore, the total elements in free list is only 3. Apart from we discussed above, we also made some other improvements/bug fixes to mheap: 1. Bug fix: macros MHEAP_ELT_OVERHEAD_BYTES & MHEAP_MIN_USER_DATA_BYTES are wrongly defined. In fact MHEAP_ELT_OVERHEAD_BYTES should be (STRUCT_OFFSET_OF (mheap_elt_t, user_data)) 2. mheap_bytes_overhead is wrongly calculating the total overhead - should be number of elements * MHEAP_ELT_OVERHEAD_BYTES 3. Do not make an element if hi_free_size is smaller than 4 times of MHEAP_MIN_USER_DATA_BYTES. This is to avoid memory fragmentation 4. Bug fix: register_node.c:336 is wrongly using vector memory, should be like this: clib_mem_is_heap_object (vec_header (r->name, 0)) 5. Bug fix: dpo_stack_from_node in dpo.c: memory leak, of parent_indices 6. Some fixes and improvements of format_mheap to show more information of heap The code including all fixes is tentatively in our private code base. It can be of course shared if wanted. Really appreciate any comments! Regards, Kingwel
