[lttng-dev] High memory consumption issue on RCU side
Mathieu Desnoyers
mathieu.desnoyers at efficios.com
Sun Sep 25 13:10:10 UTC 2016
Hi,
Did you enable the CDS_LFHT_ACCOUNTING flag for your hash tables at
creation, or only CDS_LFHT_AUTO_RESIZE ?
With only CDS_LFHT_AUTO_RESIZE, the algorithm used in check_resize()
is to verify if the current chain is longer than CHAIN_LEN_RESIZE_THRESHOLD
(which is currently 3). It effectively detects bucket collisions and
resize the hash table accordingly.
If you have both CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING flags set,
then it goes as follow: for a small table of size below
(1UL << (COUNT_COMMIT_ORDER + split_count_order)), we use the
bucket-chain-length algorithm. This is because the accounting uses
split-counters, and amortizes the cost of committing to the global
counter. So it is not precise enough for small tables.
When we are beyond the threshold, then we use the overall number of
nodes in the hash table to calculate how we should resize it.
The "resize_target" field of struct cds_lfht (in rculfhash-internal.h)
is a good way to see the number of buckets that were requested at the
last resize. This is not exposed in the public API though. You can
also try compiling rculfhash with -DDEBUG, which will enable debugging
printouts that tell you how the tables are resized. You can deduce the
number of buckets from that information.
So if you expect to have many collisions in your hash table,
I recommend you activate the CDS_LFHT_ACCOUNTING flag.
Hoping this clarifies things,
Thanks,
Mathieu
----- On Sep 24, 2016, at 2:40 PM, Evgeniy Ivanov lolkaantimat at gmail.com wrote:
> All hash tables are created with 1024 initial buckets (no limit for
> max number of buckets). Only three tables can contain at most about 5
> 000 000 nodes, the rest (I think about 5000) tables contain at most
> 1000-5000 nodes. Big tables have UUID key and CityHash, small tables
> have a complicated binary key with SuperFastHash. Binary keys are the
> same between executions, but UUID are generated on the fly and if
> there're collisions it might explain why memory footprint varies so
> much.
>
> I've set both min and max buckets limits and now RSS looks constant
> between executions. Thank you very much for pointing to this! Do I
> understand it correctly, that besides load factor rculfhash also
> resizes depending on max number of nodes in any bucket? Is there any
> way to get number of buckets (sorry if I missed it looking into API)
> allocated by table? This would help to further troubleshoot the issue.
>
>
>
> On Sat, Sep 24, 2016 at 6:34 PM, Mathieu Desnoyers
> <mathieu.desnoyers at efficios.com> wrote:
>> ----- On Sep 24, 2016, at 11:22 AM, Paul E. McKenney paulmck at linux.vnet.ibm.com
>> wrote:
>>
>>> On Sat, Sep 24, 2016 at 10:42:24AM +0300, Evgeniy Ivanov wrote:
>>>> Hi Mathieu,
>>>>
>>>> On Sat, Sep 24, 2016 at 12:59 AM, Mathieu Desnoyers
>>>> <mathieu.desnoyers at efficios.com> wrote:
>>>> > ----- On Sep 22, 2016, at 3:14 PM, Evgeniy Ivanov lolkaantimat at gmail.com wrote:
>>>> >
>>>> >> Hi all,
>>>> >>
>>>> >> I'm investigating high memory usage of my program: RSS varies between
>>>> >> executions in range 20-50 GB, though it should be determenistic. I've
>>>> >> found that all the memory is allocated in this stack:
>>>> >>
>>>> >> Allocated 17673781248 bytes in 556 allocations
>>>> >> cds_lfht_alloc_bucket_table3 from liburcu-cds.so.2.0.0
>>>> >> _do_cds_lfht_resize from liburcu-cds.so.2.0.0
>>>> >> do_resize_cb from liburcu-cds.so.2.0.0
>>>> >> call_rcu_thread from liburcu-qsbr.so.2.0.0
>>>> >> start_thread from libpthread-2.12.so
>>>> >> clone from libc-2.12.so
>>>> >>
>>>> >> According pstack it should be quiescent state. Call thread waits on syscall:
>>>> >> syscall
>>>> >> call_rcu_thread
>>>> >> start_thread
>>>> >> clone
>>>> >>
>>>> >> We use urcu-0.8.7, only rculfhash (QSBR). Is it some kind of leak in
>>>> >> RCU or any chance I misuse it? What would you recommend to
>>>> >> troubleshoot the situation?
>>>> >
>>>> > urcu-qsbr is the fastest flavor of urcu, but it is rather tricky to use well.
>>>> > Make sure that:
>>>> >
>>>> > - Each registered thread periodically reach a quiescent state, by:
>>>> > - Invoking rcu_quiescent_state periodically, and
>>>> > - Making sure to surround any blocking for relatively large amount of
>>>> > time by rcu_thread_offline()/rcu_thread_online().
>>>> >
>>>> > In urcu-qsbr, the "default" state of threads is to be within a RCU read-side.
>>>> > Therefore, if you omit any of the two advice above, you end up in a situation
>>>> > where grace periods never complete, and therefore no call_rcu() callbacks can
>>>> > be processed. This effectively acts like a big memory leak.
>>>>
>>>> It was the original assumption, but in memory stacks I don't see such
>>>> allocations for my data. Instead huge allocations happen right in
>>>> call_rcu_thread. Memory footprint for my app is about 20 GB, erasing
>>>> RCU data is a rare operation, so almost 20 GB in rcu thread looks
>>>> suspecios. I'll try to not erase any RCU protected data and reproduce
>>>> the issue (complicated thing is that under memory tracer it happens
>>>> not so often).
>>>
>>> Interesting. Trying to figure out why your call_rcu_thread() would
>>> ever allocate memory.
>>>
>>> Ah! Do your RCU callbacks allocate memory?
>>
>> In this case yes: urculfhash allocates memory within a call rcu worker
>> thread when a hash table resize is performed.
>>
>> Thanks,
>>
>> Mathieu
>>
>>>
>>> Thanx, Paul
>>>
>>>> > Hoping this helps,
>>>> >
>>>> > Thanks,
>>>> >
>>>> > Mathieu
>>>> >
>>>> >
>>>> > --
>>>> > Mathieu Desnoyers
>>>> > EfficiOS Inc.
>>>> > http://www.efficios.com
>>>>
>>>>
>>>>
>>>> --
>>>> Cheers,
>>>> Evgeniy
>>
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> http://www.efficios.com
>
>
>
> --
> Cheers,
> Evgeniy
> _______________________________________________
> lttng-dev mailing list
> lttng-dev at lists.lttng.org
> https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
More information about the lttng-dev
mailing list