Environment:
Apache Ignite 2.4 running on Amazon Linux. VM is 16CPUs/122GB ram. There is plenty of room there.
5 nodes, 12GB each
cacheMode = PARTITIONED
backups = 0
OnheapCacheEnabled = true
atomicityMode = ATOMIC
rebalacneMode = SYNC
rebalanceBatchSize = 1MB
copyOnread = false
rebalanceThrottle = 0
rebalanceThreadPoolSize = 4
Basically we have a process that populates the cache on startup and then receives periodic updates from Kafka, propagating them to the cache.
The number of elements in the cache is more or less stable over time (there is just a little fluctuation since we have a mixture of create, update and delete events), but what we have noticed is that the distribution of data across the different nodes is very uneven, with one of the nodes having at least double the number of keys (and memory utilization) as the others. Over time, that node either runs out of memory, or starts doing very long GCs and loses contact with the rest of the cluster.
My expectation was that Ignite would balance the data across the different nodes, but reality shows something completely different. Am I missing something here? Why do we see this imbalance and how do we fix it?
Thanks in advance.
Bottom line, although our hash function had good distribution, the default affinity function was not yielding a good distribution of keys (and, consequently, memory) across the nodes in the cluster. We replaced it with a very naive one (partition # % # of nodes), and that improved the distribution quite a bit (less than 2% variance).
This not a generic solution; it works for us because our entire cluster is in one VM and we don't use replication. For massive clusters crossing VM boundaries and replication, keeping the replicated data in separate servers is mandatory, and the naive approach won't cut it.
Related
I'm currently running some queries with a big gap of performance between first call (up to 2 minutes) and the following one (around 5 seconds).
This duration difference can be seen through the gremlin REST API in both execution and profile mode.
As the query is loading a big amount of data, I expect the issue is coming from the caching functionalities of Neptune in its default configuration. I was not able to find any way to improve this behavior through configuration and would be glad to have some advices in order to reduce the length of the first call.
Context :
The Neptune database is running on a db.r5.8xlarge instance, and during execution CPU always stay bellow 20%. I'm also the only user on this instance during the tests.
As we don't have differential inputs, the database is recreated on a weekly basis and switched to production once the loader has loaded everything. Our database have then a short lifetime.
The database is containing slightly above 1.000.000.000 nodes and far more edges. (probably around 10.000.000.000) Those edges are splitted across 10 types of labels, and most of them are not used in the current query.
Query :
// recordIds is a table of 50 ids.
g.V(recordIds).HasLabel("record")
// Convert local id to neptune id.
.out('local_id')
// Go to tree parent link. (either myself if edge come back, or real parent)
.bothE('tree_top_parent').inV()
// Clean duplicates.
.dedup()
// Follow the tree parent link backward to get all children, this step load a big amount of nodes members of the same tree.
.in('tree_top_parent')
.not(values('some flag').Is('Q'))
// Limitation not reached, result is between 80k and 100K nodes.
.limit(200000)
// Convert back to local id for the 80k to 100k selected nodes.
.in('local_id')
.id()
Neptune's architecture is comprised of a shared cluster "volume" (where all data is persisted and where this data is replicated 6 times across 3 availability zones) and a series of decoupled compute instances (one writer and up to 15 read replicas in a single cluster). No data is persisted on the instances however, approximately 65% of the memory capacity on an instance is reserved for a buffer pool cache. As data is read from the underlying cluster volume, it is stored in the buffer pool cache until the cache fills. Once the cache fills, a least-recently-used (LRU) eviction policy will clear buffer pool cache space for any newer reads.
It is common to see first reads be slower due to the need to fetch objects from the underlying storage. One can improve this by writing and issuing "prefetch" queries that pull in objects that they think they might need in the near future.
If you have a use case that is filling buffer pool cache and constantly seeing buffer pool cache misses (a metric one can see in the CloudWatch metrics for Neptune), then you may also want to consider using one of the "d" instance types (ex: r5d.8xlarge) and enabling the Lookup Cache feature [1]. This feature specifically focuses on improving access to property values/literals at query time by keeping them in a directly attached NVMe store on the instance.
[1] https://docs.aws.amazon.com/neptune/latest/userguide/feature-overview-lookup-cache.html
From what I've read, Riak treats all nodes in a cluster equal. However we'd like to have a heterogeneous cluster, where cpu/mem/hd is not always equal - in fact they can be very different. Each node would of course meet minimal requirements that's required for a node.
Questions:
1) What is the consequence of creating a cluster composed of machines with wildly varying specifications (cpu, diskspace, disk speed, amount and speed of memory, network speed) and
2) Can the cluster detect and compensate for such differences automatically? (assuming no here)
3) Are there ways to take care of this problem in other ways? Think of: prioritizing nodes in load balancer, based on the hardware. Something else?
I'll answer your questions, however, operating Riak in this manner is strongly discouraged as Riak assumes identical capabilities among nodes.
You could possibly have wildly varying performance characteristics
for operations against your node. In general, the "weakest node" in
the system could affect operations throughout the cluster. For
instance, during the PUT phase of an operation, a replica of the
data could be routed to the weakest node and the duration of that
operation could affect the entire PUT operation based on the PUT
operation's quorum value.
No, the cluster assumes identical hardware.
There really is no way to compensate for this.
I have a small network with computers of different hardware. Is it possible to optimize workload division between these hardware using MPI? ie. give nodes with larger ram and better cpu more data to compute? minimizing waiting time between different nodes for final reduction.
Thanks!
In my program data are divided into equal-sized batches. Each node in the network will process some of them. The result of each batch will be summed up after all batches are processed.
Can you divide the work into more batches than there are processes? If so, change your program so that instead of each process receiving one batch, the master keeps sending batches to whichever node is available, for as long as there are unassigned batches. It should be a fairly easy modification, and it will make faster nodes process more data, leading to a lower overall completion time. There are further enhancements you can make, e.g. once all batches have been assigned and a fast node is available, you could take an already assigned batch away from a slow node and reassign it to said fast node. But these may not be worth the extra effort.
If you absolutely have to work with as many batches as you have nodes, then you'll have to find some way of deciding which nodes are fast and which ones are slow. Perhaps the most robust way of doing this is to assign small, equally sized test batches to each process, and have them time their own solutions. The master can then divide the real data into appropriately sized batches for each node. The biggest downside to this approach is that if the initial speed measurement is inaccurate, then your efforts at load balancing may end up doing more harm than good. Also, depending on the exact data and algorithm you're working with, runtimes with small data sets may not be indicative of runtimes with large data sets.
Yet another way would be to take thorough measurements of each node's speed (i.e. multiple runs with large data sets) in advance, and have the master balance batch sizes according to this precompiled information. The obvious complication here is that you'll somehow have to keep this registry up to date and available.
All in all, I would recommend the very first approach: divide the work into many smaller chunks, and assign chunks to whichever node is available at the moment.
I have a filtering algorithm that needs to be applied recursively and I am not sure if MapReduce is suitable for this job. W/o giving too much away, I can say that each object that is being filtered is characterized by a collection if ordered list or queue.
The data is not huge, just about 250MB when I export from SQL to
CSV.
The mapping step is simple: the head of the list contains an object that can classify the list as belonging to one of N mapping nodes. the filtration algorithm at each node works on the collection of lists assigned to the node and at the end of the filtration, either a list remains the same as before the filtration or the head of the list is removed.
The reduce function is simple too: all the map jobs' lists are brought together and may have to be written back to disk.
When all the N nodes have returned their output, the mapping step is repeated with this new set of data.
Note: N can be as much as 2000 nodes.
Simple, but it requires perhaps up to a 1000 recursions before the algorithm's termination conditions are met.
My question is would this job be suitable for Hadoop? If not, what are my options?
The main strength of Hadoop is its ability to transparently distribute work on a large number of machines. In order to fully benefit from Hadoop your application has to be characterized, at least by the following three things:
work with large amounts of data (data which is distributed in the cluster of machines) - which would be impossible to store on one machine
be data-parallelizable (i.e. chunks of the original data can be manipulated independently from other chunks)
the problem which the application is trying to solve lends itself nicely to the MapReduce (scatter - gather) model.
It seems that out of these 3, your application has only the last 2 characteristics (with the observation that you are trying to recursively use a scatter - gather procedure - which means a large number of jobs - equal to the recursion depth; see last paragraph why this might not be appropriate for hadoop).
Given the amount of data you're trying to process, I don't see any reason why you wouldn't do it on a single machine, completely in memory. If you think you can benefit from processing that small amount of data in parallel, I would recommend focusing on multicore processing than on distributed data intensive processing. Of course, using the processing power of a networked cluster is tempting but this comes at a cost: mainly the time inefficiency given by the network communication (network being the most contended resource in a hadoop cluster) and by the I/O. In scenarios which are well-fitted to the Hadoop framework these inefficiency can be ignored because of the efficiency gained by distributing the data and the associated work on that data.
As I can see, you need 1000 jobs. The setup and the cleanup of all those jobs would be an unnecessary overhead for your scenario. Also, the overhead of network transfer is not necessary, in my opinion.
Recursive algos are hard in the distributed systems since they can lead to a quick starvation. Any middleware that would work for that needs to support distributed continuations, i.e. the ability to make a "recursive" call without holding the resources (like threads) of the calling side.
GridGain is one product that natively supports distributed continuations.
THe litmus test on distributed continuations: try to develop a naive fibonacci implementation in distributed context using recursive calls. Here's the GridGain's example that implements this using continuations.
Hope it helps.
Q&D, but I suggest you read a comparison of MongoDB and Hadoop:
http://www.osintegrators.com/whitepapers/MongoHadoopWP/index.html
Without knowing more, it's hard to tell. You might want to try both. Post your results if you do!
When configuring a Hadoop Cluster whats the scientific method to set the number of mappers/reducers for the cluster?
There is no formula. It depends on how many cores and how much memory do you have. The number of mapper + number of reducer should not exceed the number of cores in general. Keep in mind that the machine is also running Task Tracker and Data Node daemons. One of the general suggestion is more mappers than reducers. If I were you, I would run one of my typical jobs with reasonable amount of data to try it out.
Quoting from "Hadoop The Definite Guide, 3rd edition", page 306
Because MapReduce jobs are normally
I/O-bound, it makes sense to have more tasks than processors to get better
utilization.
The amount of oversubscription depends on the CPU utilization of jobs
you run, but a good rule of thumb is to have a factor of between one and two more
tasks (counting both map and reduce tasks) than processors.
A processor in the quote above is equivalent to one logical core.
But this is just in theory, and most likely each use case is different than another, some tests need to be performed. But this number can be a good start to test with.
Probably, you should also look at reducer lazy loading, which allows reducers to start later when required, so basically, number of maps slots can be increased. Don't have much idea on this though but, seems useful.
Taken from Hadoop Gyan-My blog:
No. of mappers is decided in accordance with the data locality principle as described earlier. Data Locality principle : Hadoop tries its best to run map tasks on nodes where the data is present locally to optimize on the network and inter-node communication latency. As the input data is split into pieces and fed to different map tasks, it is desirable to have all the data fed to that map task available on a single node.Since HDFS only guarantees data having size equal to its block size (64M) to be present on one node, it is advised/advocated to have the split size equal to the HDFS block size so that the map task can take advantage of this data localization. Therefore, 64M of data per mapper. If we see some mappers running for a very small period of time, try to bring down the number of mappers and make them run longer for a minute or so.
No. of reducers should be slightly less than the number of reduce slots in the cluster (the concept of slots comes in with a pre-configuration in the job/task tracker properties while configuring the cluster) so that all the reducers finish in one wave and make full utilisation of the cluster resources.