Friday, September 28, 2012

OSB, Service Callouts and OQL - Part 3



This final section of the series will focus on the corrective action to avoid Service Callout related OSB serer hangs. Before we dive into the solution, we need to briefly discus about Work Managers in WLS.

WLS Work Managers


WLS version 9 and newer releases use a concept of Work Managers (WM) and self-tuning of threads to schedule and execute server requests (internal or external). All WLS server (ExecuteThread) threads are held in a global self-tuning thread pool and request are associated with WMs. The threads in the global thread pool can grow or shrink based on some inbuilt monitoring and heuristics (grow when requests are piling up over a period of time or shrink when threads are sitting idle for long etc.) Once the request is finished, the threads go back to the global self-tuning pool. The Work Manager is a concept to associate a request with some scheduling policy and not a thread. There is a "default" WM in WLS that is automatically created. A copy or template of the "default" WM will be used for all  deployed applications, by default, out of the box. 

There are no dedicated threads associated with any WM. If an application decides to use a custom (non-default) WM, the requests meant for that application (like Webapp Servlet requests or EJB or MDB) will be scheduled based on its WM policies. The association of Work Manager with an application is via wl-dispatch-policy in the application descriptor. Multiple different applications can use a copy of Work Manager policy (each inherit the policies associated with that WM) or refer to their own custom WMs. 

Within a given Work Manager, there are options like Min-Thread and Max-Thread Constraints. Explaining the whole WM concept and the various constraints is beyond the scope of this blog entry. Please refer to Workload Management in WebLogic Server white paper on WMs and Thread Constraints in Work Managers blog post to understand more on Constraints and WMs. We will later refer to Custom WMs and Min-Thread Constraint for our particular problem. 

Suffice to say, use Custom Work Manager and Min-Thread Constraint (set to real low value, say 3 or 5) for a given application only under rare circumstances to avoid thread starvation issues like incoming requests requiring additional server thread to complete (as in case of Service callout requiring additional thread to complete the response notification) or loop-backs of requests (AppA makes outbound call which again lands on same server as new requests for AppB); using Min-Threads excessively can cause too many threads or inversion of priority as mentioned in the previously referred blog posting. Use Custom WM with Max-Thread Constraint only in case of MDBs (to increase number of MDB instances processing messages in parallel). 

Corrective Actions for handling Service Callouts


Now that we have seen (or detected) how Service Callouts can contribute to Stuck threads and thread starvation, there are solutions that can be implemented to make OSB gracefully recover from such situations.

1) Ensure the remote Backend Services invoked via Service callouts or Publish can scale under higher loads and still maintain response SLAs. Using the heap dump analysis, identify the remote services involved and improve their scalability and performance.

2) Use Route action whenever possible instead of the Service Callout when the actual invocation for a proxy is a single service and not multiple, and can be implemented using simple Route. Avoid Service callouts for calling co-located services that only do simple transformations or logging. Just invoke them as replace/insert/rename/log actions directly instead of using Service Callouts to achive the same result.

3) Protect OSB from thread starvation due to excessive usage of Service Callouts under load. The actual response handling is handled by an additional thread (Thread T2 in  the Service callout implementation image) for a very short duration and it just notifies the Proxy thread waiting on the Service callout response.

Now this is one good match for applying a custom Work Manager with Min-Thread Constraint we discussed earlier as there is a requirement for additional thread to complete a given request (a bit of loop-back). For the remote Business Service definition that is invoked via Service Callouts, we can associate a Custom Work Manager with Min-Thread constraint so the response handling part (T2) can use a thread to get scheduled right away due to the Min-Thread Constraint (as long as we have not hit the Min-Thread constraint) instead of waiting to be scheduled. Since the thread is really used just for a real short duration, its the right fit for our situation. The thread can pick the Service Callout response of the remote service from  the native Muxer layer when the response is ready and then immediately notify the waiting Service Callout thread before returning to work on the next Service Callout response or go back to the global self-tuning pool. This will ensure there are no thread starvation issues with the Service Callout pattern under high loads.

Custom Work Managers and Min-Thread Constraint


Create a custom Work Manager in WLS with a low Min Thread Constraint (less than or max 5) via the WLS Console.

Login to the WLS Console and expand the Environment node to select the Work Managers.





Start with creation of a Min Thread Constraint.
Select the Count to be 5 (or less).




Target the Constraint to the relevant servers (or OSB Cluster).
Next create a new Work Manager.

 



Target the new custom Work Manager to the relevant servers (or OSB Cluster).
Next associate the Work Manager with the previously created Min Thread Constraint.
Leave the rest as empty (Max Thread Constraint/Capacity Constraint).





Note: The Server instance would have to be restarted to pick the changes.

OSB Business Service with Custom Work Manager


Login to the sbconsole.
Create a Session
Go to the related Business Service configuration.
Edit the HTTP Transport Configuration
Select the newly created custom Work Manager for the Dispatch Policy


Save changes.




Commit the session changes

Now with the above changes to the Business Service, the OSB Service callouts will always use the custom Work Manager with Min-Thread Constraint to handle the response from the Business Service and notification of the waiting Proxy thread and not run into any of the thread starvation issues that we observed earlier with Service Callouts.

The same custom Work Manager can be used by multiple Business Services that are all invoked via Service Callout actions as the thread would be used for a very short duration and can handle responses for multiple business services. Associate the dispatch policy of the Business services invoked via Service Callouts to use the custom WM.

Summary


Hope this series gave some pointers on the internal implementation of OSB for Route Vs. Service Callouts, correct usage of Service Callouts, identifying issues with callouts using Thread Dump and Heap Dump Analysis and the solutions to resolve them. 



OSB, Service Callouts and OQL - Part 2


This section of the "OSB, Service Callouts and OQL" blog posting will delve into thread dump analysis of OSB server and detecting threading issues relating to Service Callout and using Heap Dump and OQL to identify the related Proxies and Business services involved. The previous section dealt with threading model used by OSB to handle Route and Service Callouts.

Thread dump analysis of OSB Service Callouts


There have been numerous customer situations where the performance and response times starts degrading under heavy load in OSB and users are unable to identify the cause for such slowdown and areas of bottleneck.There can be appearance of STUCK Thread notifications also. Taking multiple thread dumps at short intervals (10 - 15 seconds) from the OSB server is the first step towards identifying the problem area. Next analyze the threads for patterns - are they waiting for database response, remote invocation responses? etc.

One way to identify OSB related threads is look for com.bea.wli package name in the thread stacks. There are numerous OSB related patterns and advisories included in ThreadLogic thread dump analysis I had blogged previously. Using ThreadLogic will make the thread dump analysis part lot easier as it can parse multiple dumps as well as identify thread progress across successive dumps.

Some OSB related advisories packaged in ThreadLogic:
















From the above list, we can see ThreadLogic will attempt to identify threads involved in inbound Http Proxy, Java Callouts, Proxy in WAIT state for a response (this can be for Service Callout or Sync Publish action), Service response cache lookup (using in-built Coherence Cluster Cache), Publish action, Session activation, Web Service Callout, response handling etc and mark them with a related health level.

Keep in mind, with Route actions, even if the remote services are slow, these wont show up in thread dumps (unless it uses Exactly-Once QoS) as there is no thread actually waiting for the response and threads are only used during the request invocation and actual response processing as discussed earlier.

But for the Service Callouts, two threads needs to be used (at time of actual response hand-off). For an OSB server under heavy load and exhibiting slowness or STUCK (thread has been executing one request for more than 10 minutes in WebLogic), ThreadLogic will report something similar in the Thread Groups Summary node:

Under load, OSB related threads using service callouts will appear in following state in ThreadLogic analysis. The threads might appear as hung waiting for a response notification from the remote service endpoint.





The above snapshot shows ThreadLogic detecting and marking OSB threads involved in wait for Service Callout response (as well as Webservice callout and STUCK). The overall OSB server instance appears to show multiple threads waiting for Service response. The FATAL health level is due to threads appearing in STUCK State.

In the ThreadLogic's Merge or Diff view, one can see multiple threads executing the Service Callout, some progressing, while others remaining stuck in same code execution between dumps.



In the Merge/Diff view of ThreadLogic, the green column entries with "Progress" indicate thread progressing between dumps while the "No Change" indicates absence of thread progress. If the thread is in a bad state (WATCH or WARNING or FATAL) in a previous dump and not progressing in the next successive dump, its marked in brown background while those in IGNORE or NORMAL and not showing any progress would use yellow background.

Detecting Related Services involved using Heap Dumps


Now we have detected OSB threads which are waiting for service responses. The corrective action would be to ensure the remote service can respond in time under increasing loads along with implementing better thread management in OSB to handle the response.

Before we get into the solution implementation, we still need to identify those specific business services that are contributing to the hang situation, as the solution/remedial actions have to be applied against those specific services.

If there are very few external services that OSB invokes via Service Callout or Publish action, then we can easily identify or detect which services are contributing to the slowdown. But if there are numerous services, all invoked as mix of Service Callouts/Publish actions, then it becomes difficult to identify the related remote business services and apply the corrective actions.

The thread dumps can indicate the call pattern, but cannot provide any information on which specific Proxy or Business services were involved as the OSB framework code is generic and executes same code paths for all proxy services or Service Callouts (similar to a JDBC code stack trace that does not provide data about the SQL being executed). The OSB Console Management Dashboard returns monitored statistics for services that have completed execution, but not for those still executing or hanging. So, we would have to start analyzing web traffic patterns (access logs or web proxy logs) or network interactions (like connections to remote side) to understand incoming load but might still not get an entire picture which services are really hanging.

In such situations, analyzing the heap dump of the Server JVM instance can provide a gold mine of information.  Heap dumps contain a complete snapshot of every object instance loaded in memory and details of threads executing code or acting on objects when the heap dump is generated. Useful data can be retrieved as long as the heap dump can be analyzed by any of the commonly available Java Heap Dump Analyzers (Eclipse MAT, JVisualVM, IBM HA, Sun JHAT, YourKit etc.).

Note: There are two separate versions of Eclipse MAT - a 32 bit and 64 bit version. You cannot just switch the JVM to 64 bit and use Eclipse MAT 32 bit version to analyze big heap dumps (in excess of 4GB heap size). Best to download and use the 64 bit version of MAT.

Capturing Heap Dumps

Most jvms allow capturing heap dump from a running jvm instance. Depending on the vm version and options used, some of the dump formats might not be readable from vendor-agnostic tools like Eclipse MAT or VisualVM. Always try to capture the heap dump in HPROF format, so its not vendor specific. JRockit versions pre-R28 cannot generate heap dumps. JRockit R28+ versions allow heap dumps in hprof format.

For Sun Hotspot:
jmap is the utility (packaged within jdk bin folder) to generate heap dumps given a Java Process ID.

jmap -dump:format=b,file=heap.bin <pid>

format=b option tells to dump the heap in binary format instead of just string/text format.

For JRockit:
Use jrcmd (or JRMC) to generate heap dumps in hprof format (JRockit version should be R28+)
jrcmd <PID> hprofdump filename=<pathNameToFile>

Normally jrockit heap dumps get generated in the process's current working directory.

For IBM:
Doing a kill -3 on a IBM JVM process id generates a textual representation of the JVM process heap dump.
Please refer to IBM JVM documentation for details on generating heap dump using dump agents or programmatically (also requires some additional JVM command line arguments to dump heap in the right format).

OSB and OQL


Most users will use heap dumps mainly to detect memory leaks or usage. But there is lot more that can be researched from heap dumps using Object Query Language, referred as OQL. OQL provides SQL like syntax to navigate, parse and retrieve data from java heap dump. Navigating object relationship and class hierarchies is very simple in tools like Eclipse MAT. There are numerous OQL Tutorials and a detailed documentation available within Eclipse MAT on OQL syntax and usage.

How can we use OQL in the OSB Service Callout situation? From the thread dumps, the code pattern involved in Service callouts is "com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener" from the Stuck threads stack trace. Lets use OQL to select all instances of this class.

Sample OQL:
select * from instanceof com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener

Select the OQL option at the top (4th icon under the heap dump in MAT) and execute the OQL.

Also, as a general recommendation, always use "instanceof <type>" to select all derived types just to be safer (although there is no actual derived type from PipelineContextImpl$SynchronousListener). Eclipse MAT will execute the OQL query against the Heap Dump and return the results for all pending responses. The Heap Dump snapshot should have been captured from an OSB instance when it was actively executing Service Callout or Publish invocations.

Sample snapshot for the select com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener OQL in Eclipse MAT.



Now pick any instance from the results and the MAT Inspector tab (shows up under Window option) should provide details of the instance - member variables, references, values etc. Now the remaining steps of navigating the instances are mostly guess work - a mix of trial and errors as we don't really know the class structure or relationship between members or references, that might or might not contain useful data.

One can see there is a _service reference within the PipelineContextImpl$SynchronousListener and its of type ..BusinessServiceImpl. Lets drill into it. Expanding it shows another member reference "_ref".


Clicking on the "_ref" shows more interesting data about the "_service". It shows the actual Business Service name and path in its "fullname" attribute. Now we know which services are being invoked via Service Callout or Publish action.

Use OQL and the navigation logic to get list the Business services invoked that are still pending for response.

OQL:  
SELECT toString(s._service._ref.fullname) FROM INSTANCEOF com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener s 

Note: Use toString(variable) to get the actual String content instead of a reference to the String in Eclipse MAT. Use ref.toString() in JVisualVM. Other primitives will be reported directly.


There is also a "this$0" reference for the PipelineContextImpl$SynchronousListener. Drilling into it yields "_proxy". This is the actual reference to the calling proxy. Drill into "_proxy" shows its own "_ref" member attribute. The "fullname" of the "_ref" provides the name of the Proxy that was invoking the Service Callout.



So, now we have traversed from instance of PipelineContextImpl$SynchronousListener to "this$0" to "_proxy" to "_ref" to "fullname" to arrive at Proxy name. Similarly, we navigated PipelineContextImpl$SynchronousListener instance to "_service" to "_ref" to "fullname" to identify the Business Service. Now lets run a single OQL to get both the calling Proxy and the Business Service executed via Service Callout or Publish action that blocks for response.

OQL:
SELECT toString(s.this$0._proxy._ref.fullname) AS ProxyService, toString(s._service._ref.fullname) AS BusinessService FROM INSTANCEOF com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener s 

The sample output should appear like:











To analyze the same heap dump on JVisualVM (under hotspot jdk/bin folder), use following OQL:
select s.this$0._proxy._ref.fullname.toString() , s._service._ref.fullname.toString()  from com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener s 




The OQL syntax (select/from/instanceof are all case sensitive in JVisualVM), columns display and navigation/drill down are much more easier with Eclipse MAT compared to JVisualVM.

Lets run a more detailed OQL to report on type of remote invocation - whether its a Service Callout or Publish action. If there is a response Handler for the proxy, then its a Service Callout; if null, its a Publish action.

OQL:
SELECT toString(s.this$0._proxy._ref.fullname) AS InvokingProxy, toString(s._service._ref.fullname) AS OutboundService, s._responseHandler AS PublishOrWSCallout FROM com.bea.wli.sb.pipeline.PipelineContextImpl$SynchronousListener s 




We have now analyzed the call patterns for the OSB Server threads blocked for remote response and the remote services involved using OQL on the Heap Dump file. Next would be to focus on the correction action. Before that, we need to briefly discus about Work Managers in WLS.

Summary


Hope this section in the "OSB, Service Callouts and OQL" series gave some pointers on identifying issues with Service Callouts using Thread Dump and Heap Dump Analysis. The next section will go more into WebLogic Server Work Manager concept and the corrective actions to solve OSB Service Callout related hangs.






OSB, Service Callouts and OQL - Part 1

Oracle Fusion Middleware customers use Oracle Service Bus (OSB) for virtualizing Service endpoints and implementing stateless service orchestrations. Behind the performance and speed of OSB, there are a couple of key design implementations that can affect application performance and behavior under heavy load. One of the heavily used feature in OSB is the Service Callout pipeline action for message enrichment and invoking multiple services as part of one single orchestration. Overuse of this feature, without understanding its internal implementation, can lead to serious problems.

This post will delve into OSB internals, the problem associated with usage of Service Callout under high loads, diagnosing it via thread dump and heap dump analysis using tools like ThreadLogic and OQL (Object Query Language) and resolving it. The first section in the series will mainly cover the threading model used internally by OSB for implementing Route Vs. Service Callouts.

OSB Pipeline actions for Service Invocations


A Proxy is the inbound portion of OSB that can handle the incoming request, transform/validate/enrich/manipulate the payload before invoking co-located or remote services. The execution logic is built using the proxy pipeline actions. For executing the remote (or even local) business service, OSB provides three forms of service invocations within a Proxy pipeline:
  • Route - invoke a single business service endpoint with (or without) a response. This happens entirely at end of a proxy service pipeline execution and bridges the request and response pipeline. The route can be treated as the logical destination to reach or final service invocation. There can be only one Route action (there can be choices of Route actions - but only one actual execution) in a given Proxy execution.
  • Publish - invoke a business service without waiting for result or response (like 1-way). The caller does not care much about the response. Just interested in sending out something (and ensuring it reaches the other side).
  • Service Callout - invoke one or more business service(s) as part of message augmentation or enrichment or validation but this is not the primary business service for a given Proxy, unlike the Route action. The service callouts can be equivalent to credit card validation, address verification while Route is equivalent to final order placement. There can be multiple Service Callouts inside a Proxy pipeline.


OSB Route Action

Most HTTP remote service invocations with responses are synchronous and blocking in nature. The caller creates a payload, connects to the business service endpoint, transmits the payload and waits for a response. The caller has to wait till the response is ready and transmitted back. Using Java Native IO, one can avoid the blocking wait for response and only read the response once its ready. But this is not an easy option for higher level applications that aim at SOAP, XML, REST forms of service interactions. They need threads to wait for the response and if the remote business services are slow, more threads can get tied up instead of working on other tasks.

When using the Route Action for HTTP based Business services, OSB does not tie up a thread waiting for the remote response. Instead it leverages Native IO within WebLogic Server Muxer Layer and Future Response AsyncServlet functionality to decouple the caller thread from the actual response handling portion thereby behaving asynchronously. When the client makes a request to OSB Proxy and the request pipeline finally executes a Route action, OSB posts the request one-way and registers a future Response Async Servlet method to receive a callback of the response.



The proxy thread that processed the request pipeline path makes the outbound call and returns, without waiting for the actual response. This thread is then free to execute other pending requests. The WLS Muxer layer detects when there is response data readily available to be read from the socket for that outbound business service call and then triggers a callback to the OSB's registered Async Servlet. Now a different thread picks the response and then execute the response pipeline flow within OSB. This way, the proxy uses two threads for segregating the request from the response processing in the Route Action. This translates to OSB using minimal threads for service executions, without blocking for response, even if the remote service is slow. But for the external client calling into OSB, it appears like one synchronous blocking call, while OSB keeps its thread usage to the minimum and handles more requests, without using additional threads or waiting for remote service responses.

By default, for most HTTP based interactions for both incoming Proxy service and external Route, there would be no transactions involved and so the Route action would use Best-Effort QoS (Quality of Service) and would leverage the async threading model described previously. However, if the Route is invoked as part of an existing transaction (if the calling Proxy service was JMS with XA Connection Factory enabled or other Transactional proxy service invocation like Tuxedo or started off a Transaction in the middle of the pipeline) and wants to use Exactly-Once QoS, then the invoking Thread (T1) of the Route action blocks till a response is received and then commits the transaction. The response is only then picked by another thread  (T2) after the Route action is completely successful and transaction committed. So, the thread invoking Route will appear as blocking. If the QoS is changed to Best-Effort, then the async threading model will be used as in case of HTTP based service invocations.


OSB Service Callout Action

A Service Callout is not the actual target or end service for a Proxy Service in OSB. Its simply a service invocation to either modify, validate, transform, augment or enrich the incoming request or outgoing response within a proxy execution. It can be invoked from either the request or response path. Multiple Service Callouts can also be executed in any order or fashion. Route is the final target and so there can only be one route in a proxy execution. Service callouts are used when a response is needed from the service execution. So, the caller of the Service Callout will block till a response is available. If responses are not needed or its strictly one-way sends, Publish Action can be used.

Most users will consider the OSB Service Callout to be similar to Route action. Both are invoking some remote service and ultimately getting back some response. The caller of the proxy blocks till the response is received. The time used by the remote service in sending back the response cannot be cut down from the final proxy response time. But the request and response handling part differs considerably in the Service Callout compared to the Route Action.


Unlike in Route where the invoking thread returns right away after making the remote invocation, the Service Callout thread T1 actually waits for a notification of response for that invocation; it does not really handle the response directly from the remote service. When the remote service sends back a response, the WLS native Muxer layer picks it and then schedules another thread T2 to handle it. The thread T2 does not really do much other than notify the waiting T1 thread of the availability of response and return. Now T1 wakes up from its waiting state and then continues execution of the rest of the proxy pipeline logic. So, in case of Service Callout, the original thread T1 actually waits for the response to become available, while another thread T2 is needed to pick the response and then notifies T1. So, essentially two threads will be used with one thread (T1) completed dedicated for duration of the service callout and beyond and another thread for a short while. In Route, threads T1 and T2 are never used concurrently and also, are not wasted or needed, when the response is yet to be sent across from the remote service.

This design implementation of Service Callout action can affect the behavior of OSB under high loads when there is heavy use of Service Callouts to either aggregate data from multiple services or just used repeatedly for VERT (Validate, Enrich, Route, Transform) messages instead of using Routing action. As more requests repeatedly use Service callouts, these can tie up valuable threads waiting for the response from remote or other local services while there are no more threads available to handle the actual incoming response and notify the waiting Service callout threads. In summary, overuse of Service callouts can lead to thread starvation issues and degraded performance under heavy loads.

For a synchronous publish (like Exactly-Once QoS Publish) that has to wait for confirmation and response, the behavior is the same as in Service Callout - requires two threads for the waiting and notification.

Summary


Hope this post gave some pointers on the internal implementation of OSB for Route Vs. Service Callouts and correct usage of Service Callouts. The remaining sections will deal with identifying issues with callouts using Thread Dump and Heap Dump Analysis and the corrective actions to resolve them.