Rana's work shows a clear user requirement (@Xikui pay attention :-)) --
we need two forms of parallelism hint, one that does what we currently
do - which is widen the parallelism AFTER reading from storage at the
first opportunity to do so - and another that widens it IMMEDIATELY
(somehow :-)). The latter is clearly what Rana would ideally have been
able to make use of, so she wouldn't have to change the data layout to
get more parallelism. Food for thought.
On 1/28/18 8:29 PM, Murtadha Hubail wrote:
If reloading the data isn’t too much trouble, the first thing I would
do is recreate the instance with more partitions (e.g. partition per
core or partition per 2 cores) and check the cores utilization. If
this is the same dataset as the one in your previous email, you
mentioned that it was about 10GB per partition, in that case, you
might want to allocate at least 40GB for the buffer cache and you can
reduce storage.memorycomponent.globalbudget to get enough memory to
execute the job (depending on the number of partitions you create).
After recreating with higher number of partitions, don’t use “SET
`compiler.parallelism` "39"”. It will automatically use the number of
partitions you create.
Regarding the metrics time, it includes the results printing time, so
if you want to see if it has any impact, try adding “limit 1” at the
end of your query or change it to select count(*) instead of subject_id.
Cheers,
Murtadha
*From: *Rana Alotaibi <[email protected]>
*Date: *Monday, 29 January 2018 at 6:48 AM
*To: *<[email protected]>
*Cc: *<[email protected]>, <[email protected]>
*Subject: *Re: Hyracks Job Requirement Configuration
*- Do you see all cores being fully utilized during the query execution? *
**I have noticed only 6 cores were utilized
*- How much time does the query take right now and how do you measure
the query execution time? Do you wait for the result to be printed
somewhere (e.g. in the browser)?*
I'm using the HTTP APIs. The response is a JSON object that includes
the query execution time:
{ "status": "success",
"metrics": {
*"elapsedTime": "434.627299814s",
"executionTime": "434.626137977s",*
"resultCount": 4943,
"resultSize": 132293,-
"processedObjects": 46875
}
}
I run the query 10 times and took the average which is ~6mins.
*- You mentioned that you have 4 partitions, how many physical hard
drives are they mapped to?*
**One physical hard drive
*- Also, increasing the sort/join memory doesn’t necessarily lead to a
better performance. Have you tried changing these values to something
smaller and seeing the effects?*
Yes, I tried the following numbers:
1) sort-memory: 32MB, join-memory: 64MB
2) sort-memory: 64MB, join-memory: 128MB
3) sort-memory: 128MB, join-memory: 265MB
The execution time remains on average ~6 - 6.5mins. I didn't see any
improvement. The configurations that I have now:
- compiler.parallelism :39 //Only 6 were utilized
- storage.buffercache.size: 20GB
- storage.buffercache.pagesize: 1MB
Thanks,
Rana
On Sun, Jan 28, 2018 at 6:41 PM, Murtadha Hubail
<[email protected]<mailto:[email protected]>> wrote:
I have few questions if you don’t mind:
Do you see all cores being fully utilized during the query execution?
How much time does the query take right now and how do you measure
the query execution time? Do you wait for the result to be printed
somewhere (e.g. in the browser)?
You mentioned that you have 4 partitions, how many physical hard
drives are they mapped to?
Also, increasing the sort/join memory doesn’t necessarily lead to
a better performance. Have you tried changing these values to
something smaller and seeing the effects?
Cheers,
Murtadha
*From: *Rana Alotaibi
<[email protected]<mailto:[email protected]>>
*Date: *Monday, 29 January 2018 at 5:21 AM
*To: *<[email protected]<mailto:[email protected]>>
*Cc:
*<[email protected]<mailto:[email protected]>>,
<[email protected]<mailto:[email protected]>>
*Subject: *Re: Hyracks Job Requirement Configuration
Thanks Murtadha! The problem solved. However, increasing the
number of cores didn't help to improve the performance of that query.
On Sun, Jan 28, 2018 at 5:05 PM, Murtadha Hubail
<[email protected]<mailto:[email protected]>> wrote:
Hi Rana,
The memory used for query processing is automatically
calculated as follows:
JVM Max Memory - storage.buffercache.size -
storage.memorycomponent.globalbudget
The documentation defaults for these parameters are outdated.
The default value for storage.buffercache.size is (JVM Max
Memory / 4) and it's the same for
storage.memorycomponent.globalbudget. Since your dataset is
already loaded, you could reduce the budget of
storage.memorycomponent.globalbudget. In addition, if I recall
correctly, your dataset size is way smaller than what's
allocated for the buffer cache, so you might want to reduce
the buffer cache budget. That should give you more than enough
memory to execute on 39 cores.
Cheers,
Murtadha
On 01/29/2018, 3:30 AM, "Mike Carey"
<[email protected]<mailto:[email protected]>> wrote:
+ dev
On 1/28/18 3:37 PM, Rana Alotaibi wrote:
> Hi all,
>
> I would like to make AsterixDB utilizes all available
CPU cores (39)
> that I have for the following query:
>
> USE mimiciii;
> SET `compiler.parallelism` "39";
> SET `compiler.sortmemory` "128MB";
> SET `compiler.joinmemory` "265MB";
> SELECT P.SUBJECT_ID
> FROM LABITEMS I, PATIENTS P, P.ADMISSIONS A, A.LABEVENTS E
> WHERE E.ITEMID/*+bcast*/=I.ITEMID AND
> E.FLAG = 'abnormal' AND
> I.FLUID='Blood' AND
> I.LABEL='Haptoglobin'
>
>
> The total memory size that I have is 125GB(57GB for the
AsterixDB
> buffer cache). By running the above query, I got the
following error:
>
> "msg": "HYR0009: Job requirement (memory: 10705403904
bytes, CPU
> cores: 39) exceeds capacity (memory:
3258744832<tel:%28325%29%20874-4832>bytes, CPU cores: 39)"
>
> How can I change this capacity default configuration?
I'm looking into
> this page :
https://asterixdb.apache.org/docs/0.9.2/ncservice.html.
> Could you please point me to the appropriate
configuration parameter?
>
> Thanks
> -- Rana
>
>
>
>
