spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From David Capwell <dcapw...@gmail.com>
Subject Re: SizeEstimator
Date Tue, 27 Feb 2018 03:46:19 GMT
advancedxy <advancedxy@gmail.com>, I don't remember the code as well
anymore but what we hit was a very simple schema (string, long). The issue
is the buffer had a million of these so SizeEstimator of the buffer had to
keep recalculating the same elements over and over again.  SizeEstimator
was on-cpu about 30% of the time, bounding the buffer got it to be < 5%
(going off memory so may be off).

The cost is really (assuming memory is O(1) which is not true) O(N × M)
where N is number of rows in buffer and M is size of schema.  My case could
be solved by not recomputing which would bring the cost to O(M) since
bookkeeping should be consistent time. There was logic to delay
recalculating bases off a change in frequency, but that didn't really do
much for us, bounding and spilling was the bigger win in our case.

On Mon, Feb 26, 2018, 7:24 PM Xin Liu <xin.e.liu@gmail.com> wrote:

> Thanks David. Another solution is to convert the protobuf object to byte
> array, It does speed up SizeEstimator
>
> On Mon, Feb 26, 2018 at 5:34 PM, David Capwell <dcapwell@gmail.com> wrote:
>
>> This is used to predict the current cost of memory so spark knows to
>> flush or not. This is very costly for us so we use a flag marked in the
>> code as private to lower the cost
>>
>> spark.shuffle.spill.numElementsForceSpillThreshold (on phone hope no
>> typo) - how many records before flush
>>
>> This lowers the cost because it let's us leave data in young, if we don't
>> bound we get everyone promoted to old and GC becomes a issue.  This doesn't
>> solve the fact that the walk is slow, but lowers the cost of GC. For us we
>> make sure to have spare memory on the system for page cache so spilling to
>> disk for us is a memory write 99% of the time.  If your host has less free
>> memory spilling may become more expensive.
>>
>>
>> If the walk is your bottleneck and not GC then I would recommend JOL and
>> guessing to better predict memory.
>>
>> On Mon, Feb 26, 2018, 4:47 PM Xin Liu <xin.e.liu@gmail.com> wrote:
>>
>>> Hi folks,
>>>
>>> We have a situation where, shuffled data is protobuf based, and
>>> SizeEstimator is taking a lot of time.
>>>
>>> We have tried to override SizeEstimator to return a constant value,
>>> which speeds up things a lot.
>>>
>>> My questions, what is the side effect of disabling SizeEstimator? Is it
>>> just spark do memory reallocation, or there is more severe consequences?
>>>
>>> Thanks!
>>>
>>
>

Mime
View raw message