[
https://issues.apache.org/jira/browse/SPARK-13718?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15193287#comment-15193287
]
Ioannis Deligiannis commented on SPARK-13718:
---------------------------------------------
(reply per paragraph)
That is the point I am trying to get through. An RDD operation will access all
RDDs but the processing weight is not the same for every data partition, which
is why I gave the fibo example. Since, scheduling is bound to the RDD partition
locality it under performs in many occasions. What you describe stands true
when data and processing is even which is not the case in most
real-time(=interactive) applications as requests as user driven and this looks
more like a normal distribution (so you will get hot-partitions).
Let try and explain how this translates. My cached RDD is 80GB containing a
billion records. One(out of thousands) specific partition contains a value say
'X'. In pseudo-code the logic looks like: rdd.filter(f=='X').map("very
expensive op").reduce(...). So user queries will quickly go over the cached
partitions, not really do anything except in the case of 'X'. So, what we are
suggesting here is that using an extra 80GB is logical, practical or cost
effective? Is there a way to actually make the attached example really scale?
(Note that the cluster is not utilized by external processes but by this code
itself).
The way I can imagine this working add this on the RDD level. Similarly to
`persist(MEMORY_ONLY)`, attached caching policies. These could be interfaced so
even a user can attach an eviction policy to a RDD. Out of the box, the basics
would be nice (LRU,RR,MRU,LFU) which could also apply to the way Sparks
currently evicts partitions. In terms of replicas, an API like
`rdd.setMaxDynamicReplication(int x)` would also help optimize scheduling.
Finally, for cases where the cached RDD creations is cheap/simple, we could use
something like "rdd.onNonLocality(Fetch/ReCreate) to either fetch from
memory(way it works now when RDD is cached) or recreate (way it works now when
RDD is not cached and replicas > 1). Note that the proposed names and available
options are not suggestions but try to set the basis on how this could work.
> Scheduler "creating" straggler node
> ------------------------------------
>
> Key: SPARK-13718
> URL: https://issues.apache.org/jira/browse/SPARK-13718
> Project: Spark
> Issue Type: Improvement
> Components: Scheduler, Spark Core
> Affects Versions: 1.3.1
> Environment: Spark 1.3.1
> MapR-FS
> Single Rack
> Standalone mode scheduling
> 8 node cluster
> 48 cores & 512 RAM / node
> Data Replication factor of 3
> Each Node has 4 Spark executors configured with 12 cores each and 22GB of RAM.
> Reporter: Ioannis Deligiannis
> Priority: Minor
> Attachments: TestIssue.java, spark_struggler.jpg
>
>
> *Data:*
> * Assume an even distribution of data across the cluster with a replication
> factor of 3.
> * In-memory data are partitioned in 128 chunks (384 cores in total, i.e. 3
> requests can be executed concurrently(-ish) )
> *Action:*
> * Action is a simple sequence of map/filter/reduce.
> * The action operates upon and returns a small subset of data (following the
> full map over the data).
> * Data are 1 x cached serialized in memory (Kryo), so calling the action
> should not hit the disk under normal conditions.
> * Action network usage is low as it returns a small number of aggregated
> results and does not require excessive shuffling
> * Under low or moderate load, each action is expected to complete in less
> than 2 seconds
> *H/W Outlook*
> When the action is called in high numbers, initially the cluster CPU gets
> close to 100% (which is expected & intended).
> After a while, the cluster utilization reduces significantly with only one
> (struggler) node having 100% CPU and fully utilized network.
> *Diagnosis:*
> 1. Attached a profiler to the driver and executors to monitor GC or I/O
> issues and everything is normal under low or heavy usage.
> 2. Cluster network usage is very low
> 3. No issues on Spark UI except that tasks begin to move from LOCAL to ANY
> *Cause (Corrected as found details in code):*
> 1. Node 'H' is doing marginally more work than the rest (being a little
> slower and at almost 100% CPU)
> 2. Scheduler hits the default 3000 millis spark.locality.wait and assigns the
> task to other nodes
> 3. One of the nodes 'X' that accepted the task will try to access the data
> from node 'H' HDD. This adds Network I/O to node and also some extra CPU for
> I/O.
> 4. 'X' time to complete increases ~5x as it goes over Network
> 5. Eventually, every node will have a task that is waiting to fetch that
> specific partition from node 'H' so cluster is basically blocked on a single
> node
> What I managed to figure out from the code is that this is because if an RDD
> is cached, it will make use of BlockManager.getRemote() and will not
> recompute the DAG part that resulted in this RDD and hence always hit the
> node that has cached the RDD.
> * Proposed Fix *
> I have not worked with Scala & Spark source code enough to propose a code
> fix, but on a high level, when a task hits the 'spark.locality.wait' timeout,
> it could make use of a new configuration e.g.
> recomputeRddAfterLocalityTimeout instead of always trying to get the cached
> RDD. This would be very useful if it could also be manually set on the RDD.
> *Workaround*
> Playing with 'spark.locality.wait' values, there is a deterministic value
> depending on partitions and config where the problem ceases to exist.
> *PS1* : Don't have enough Scala skils to follow the issue or propose a fix,
> but I hope that this has enough information to make sense.
> *PS2* : Debugging this issue made me realize that there can be a lot of
> use-cases that trigger this behaviour
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
