Hi Dave,
The perf backtrace (taken from "control-only" lcore 0) is as follows:
- 91.87% vpp_main libvppinfra.so.0.0.0 [.] mheap_get_aligned
- mheap_get_aligned
- 99.48% map_add_del_psid
vl_api_map_add_del_rule_t_handler
vl_msg_api_handler_with_vm_node
memclnt_process
vlib_process_bootstrap
clib_calljmp
Using DPDK's rte_malloc_socket(), CPU consumption drops to around 0,5%.
>From my (somewhat brief) mheap code analysis, it looks like mheap might not
take into account alignment when looking for free space to allocate
structure. So, in my case, when I keep allocating 16B objects with 64B
alignment, it starts to examine each hole it left by previous object's
allocation alignment and only then realize it cannot be used because of
alignment. But of course I might be wrong and the root cause is entirely
elsewhere...
In my test, I'm just adding 300,000 tunnels (one domain+one rule).
Unfortunately, rte_malloc() provides only aligned memory allocation, not
aligned-at-offset. Theoretically we could provide wrapper around it, but
that would need some careful coding and a lot of testing. I made an attempt
to quickly replace mheap globally, but of course it ended up in utter
failure.
Right now, I added a concept of external allocator to clib (via function
pointers), I'm enabling it only upon DPDK plugin initialization. However,
such approach requires using it directly instead of clib alloc, (e.g. I did
it upon rule adding). While it does not add dependency on DPDK, I'm not
fully satisfied, because it would need manual replacement of all allocation
calls. If you want, I can share the patch.
Best Regards,
Jacek.
2017-09-05 15:30 GMT+02:00 Dave Barach (dbarach) <dbar...@cisco.com>:
> Dear Jacek,
>
>
>
> Use of the clib memory allocator is mainly historical. It’s elegant in a
> couple of ways - including built-in leak-finding - but it has been known to
> backfire in terms of performance. Individual mheaps are limited to 4gb in a
> [typical] 32-bit vector length image.
>
>
>
> Note that the idiosyncratic mheap API functions “tell me how long this
> object really is” and “allocate N bytes aligned to a boundary at a certain
> offset” are used all over the place.
>
>
>
> I wouldn’t mind replacing it - so long as we don’t create a hard
> dependency on the dpdk - but before we go there...: Tell me a bit about the
> scenario at hand. What are we repeatedly allocating / freeing? That’s
> almost never necessary...
>
>
>
> Can you easily share the offending backtrace?
>
>
>
> Thanks… Dave
>
>
>
> *From:* vpp-dev-boun...@lists.fd.io [mailto:vpp-dev-boun...@lists.fd.io] *On
> Behalf Of *Jacek Siuda
> *Sent:* Tuesday, September 5, 2017 9:08 AM
> *To:* vpp-dev@lists.fd.io
> *Subject:* [vpp-dev] mheap performance
>
>
>
> Hi,
>
> I'm conducting a tunnel test using VPP (vnet) map with the following
> parameters:
>
> ea_bits_len=0, psid_offset=16, psid=length, single rule for each domain;
> total number of tunnels: 300000, total number of control messages: 600k.
>
> My problem is with simple adding tunnels. After adding more than
> ~150k-200k, performance drops significantly: first 100k is added in ~3s (on
> asynchronous C client), next 100k in another ~5s, but the last 100k takes
> ~37s to add; in total: ~45s. Python clients are performing even worse: 32
> minutes(!) for 300k tunnels with synchronous (blocking) version and ~95s
> with asynchronous. The python clients are expected to perform a bit worse
> according to vpp docs, but I was worried by non-linear time of single
> tunnel addition that is visible even on C client.
>
> While investigating this using perf, I found the culprit: it is the memory
> allocation done for ip address by rule addition request.
> The memory is allocated by clib, which is using mheap library (~98% of cpu
> consumption). I looked into mheap and it looks a bit complicated for
> allocating a short object.
> I've done a short experiment by replacing (in vnet/map/ only) clib
> allocation with DPDK rte_malloc() and achieved a way better performance:
> 300k tunnels in ~5-6s with the same C-client, and respectively ~70s and
> ~30-40s with Python clients. Also, I haven't noticed any negative impact on
> packet throughput with my experimental allocator.
>
> So, here are my questions:
>
> 1) Did someone other reported performance penalties for using mheap
> library? I've searched the list archive and could not find any related
> questions.
>
> 2) Why mheap library was chosen to be used in clib? Are there any
> performance benefits in some scenarios?
>
> 3) Are there any (long- or short-term) plans to replace memory management
> in clib with some other library?
>
> 4) I wonder, if I'd like to upstream my solution, how should I approach
> customization of memory allocation, so it would be accepted by community.
> Installable function pointers defaulting to clib?
>
>
>
> Best Regards,
>
> Jacek Siuda.
>
>
>
>
>
_______________________________________________
vpp-dev mailing list
vpp-dev@lists.fd.io
https://lists.fd.io/mailman/listinfo/vpp-dev
