Hello Spark community, We have a project where we want to use Spark as computation engine to perform calculations and return result via REST services. Working with Spark we have learned how to do things to make it work faster and finally optimize our code to produce results in acceptable time (1-2 seconds). But when we tried to test it under concurrent load we realizes that time grows significantly while increasing amount of concurrent requests. This was expected and we are trying to find the way to scale our solution to get acceptable time under concurrent load, and here we faced with fact that adding more slave servers not increasing average timing while having several concurrent requests.
As for now I observing following behavior: While hitting our test REST service using 100 threads and having 1 master and one slave node we have average timing for those 100 requests 30.6 seconds/request, in case we add 2 slave nodes average time becomes 29.8 seconds/request, which seems pretty similar to test case with one slave node. While doing those test cases we monitor server load using htop and weird thing here that in first case our slave node’s CPU’s were loaded on 90-95% and in second case with 2 slaves it loads CPU’s on 45-50%. We are trying to find bottleneck in our solution but was not succeed in this exercise yet. We have checked all hardware for possible bottleneck Network IO, Disk IO, RAM, CPU but no-one seems to be even closed to limit. Our major suspects at this moment is Spark configuration. Our application is Self Contained Spark application, Spark is ruined in standalone mode (without external resource managers). We are not submitting jar to spark using shell script, instead we create spark context in our spring boot application and it connects to spark master and slave by itself. All requests to spark are going through internal thread pool. We have experimented with thread pool sizes and find out the best performance appears when we have 16 threads in thread pool, where each thread is performing one or several manipulations with RDDs, and by itself could be considered as spark job. For now I assume we either misconfigured Spark due to lack of experience with it, or our use case is not really use case for Spark and it is simply not designed for big parallel load. I’ve make this conclusion since behavior when after adding new nodes we have no performance gain doesn’t makes any sense for me. So any help and ideas would be really helpful. Hardware details: Azure D5v2 instances for master and 3 slaves. D5v2 featured with 2.4 GHz E5-2673 v3 (16 cores, 56Gb RAM, SSD HDD). Using ipref we have tested network speed and it is around 1 Gb/s. -- CONFIDENTIALITY NOTICE: This email and files attached to it are confidential. If you are not the intended recipient you are hereby notified that using, copying, distributing or taking any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error please notify the sender and delete this email. --------------------------------------------------------------------- To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org