Why performance matters #1

One of the many issues that i've faced as an enterprise java developer has been performance. It is important to understand the issues surrounding performance and address them before the problems escalate into unresolvable issues. It is probably the fixation of architectural design patterns and conventions with the enterprise java community that lead us to overlook various performance related issues.

The community has today moved a long way from MVC frameworks and monolithic j2ee 1.4 servers to increasing adoption of various open source frameworks and protals. The enterprise applications too have evolved from websites having component based solutions to portals and coarse granular SOA applications, built on top of web services. However, one thing remains constant, that is the layered execution model of any application. This can be demonstrated as follows :

Our Java EE Application
Application Framework
Application Server
Java Runtime Environment
Operating System
Hardware

It is important to note that the performance of the solution does'nt depends on the application alone, rather, it depends on correct performance tuning and optimization of all the layers associated with the application. This form of issue is known as Vertical complexity.

In actual applications, however, there are a lot of discrete components that are present, each with its own 'stack' of complexity. Together, this forms into another complexity known as horizontal complexity. These issues are apparant only when the application faces heavy load in its lifecycle or when it begins to operate beyond a single JVM. This can result in the following problems :-

• Slow execution of application- beyond the aggreed upon threshold in SLAs
• Application performance degradation over a period of time- memory leaks, resource allocation bugs
• Erratic CPU utilization and application freezing
• Performance anomalies which occur in production- hard to reproduce, which escape load testing

So, it is not hard to imagine why organizations and teams look forward to solve the problems. Solution to these problems has also a fancy name, “Application Performance Management” which is a set of recepies and approaches meant to address the issues listed above. In development, this is applied to memory analysis, code profiling and test coverage. Memory analysis can be done using any standard debugger and most IDEs usually have a nice interface to do so. Profiling, however involves some serious investment into the profiler that you are planning to use. Fortunately, some IDEs like netbeans have in built profilers, but external profilers like yourkit are more relied upon. A code coverage tool does the same thing for performance testing as it normally does for unit testing- identifying whether a given code is covered under performace testing or not.

Returning to our 'stack' for a moment, we can easily visualize the tasks needed for performance optimization at different levels.

While the performance at application level can be done via APM, the underlying application framework also plays an important role. For instance, if the framework is in beta and the application is using some new feature of it, chances are that it might have a bug that can wreak havoc upon the application and create confusing problems. The application server also needs to be properly tuned and customized for clustering and performance scaling if needed. In other cases too, there is a need for system administration of these servers as they not only act as the host for the application, but also provide monitoring activities (which can further be used as performance measurement) and manage external resources needed by the application. JVM is an often ovelooked slice of the stack, which if used correctly, can cause precise Garbage Collection trips, memory management for the application as well as its dependencies. It never hurts to have knowledge of the JVM memory areas like stack, heap and perm spaces, which are the places where the application and its server reside and operate. The operating system and hardware are the obvious choices for performance optimization because these provide the underlying resource allocation, scalability, computation and other aspects that are needed for the software. From a developer's point of view, the operating system and hardware is not generally modifiable, but the overlying layers are. So it is better to have an understanding of these concepts. In a future post, i'll start with analyzing specific performance issues relating to a tier and post recepies and solutions that I came up with so far.

Scalable Application Architecture with Memory Caches

This post is dedicated towards an important feature that is required in creation of highly performant and scalable applications. Data has traditionally been stored in specialized software, aptly known as database. However, given the massive use of some specific data which is generally static in nature in applications that permit large number of users(thousands, if not more) to interact with the application simultaneously. Using a memory storage instead of a hard disk file results in huge performance gain as the requests pertaining to that data would not be read from a database located on a disk.Think of a in-memory data cache as a robust, persistent storage that can be used in certain transient tasks where the data is stored in RAM of cache machines.

The caching mechanism (framework responsible for maintaining the cache) has to do the following main tasks :

• Maintain a cache (Obvious one)
• Determine the pattern of requests (which requests are more in number)
• Flush out resources and load new ones (different strategies can be used here, most common one is LRU)
• Make sure that the data maintained in cache is correct (and if not, then to which extent)
• Enforce consistency of cache across different machines (Generally this design is needed in cloud or clustered machines)
• Maintain resource utilization (Evict the data from cache if server need increases)

Caching is generally found in distributed applications which are targeted to be used by large amount of people if they are not already in production environment. Today, most of the memory cache frameworks do not offer a synchronization mechanism between the data stored in the database and the memory cache. To overcome this problem, we need to explicitly set an expiration value of the cached object so that it gets refreshed upon requests after a certain period. This performance optimization can be done not only for database specific operations, but also on other data such as repeated web service calls, computation results, static content, etc.

A popular interface standard for java is the JCache, which was proposed as JSR 107. A memory cache software, eg: memcatched [http://www.memcached.org], stores key value pair of data. As soon as a request is generated, the key values are searched, which results in a cache hit or a cache miss scenario. JSR 107 has been adopted in different implementations, one of which is the Google App Engine, which is a cloud platform supporting python and java runtimes. Here, this comes in form of a memcache service for the java runtime. This can be better explained with the following example :


import com.google.appengine.api.memcache.MemcacheService;
import com.google.appengine.api.memcache.MemcacheServiceFactory;
......

MemcacheService cache=MemcacheServiceFactory.getMemcacheService();

.....

cache.put("key","value");

....

object=cache.get("key");

.....
cache.delete("key");

This really makes application scalability easier for the developers (Cloud environments do induce the responsibility of creating applications that can scale quickly). As of now, similar feature doesnt exist in Windows Azure, but what the future holds for this technology cannot be speculated.
Thus it is not surprising that this technology is used in prime websites like YouTube, Wikipedia, Amazon,SourceForge, Metacafe, Facebook, Twitter, etc and that too in large quantites (for eg: Facebook uses over 25 TB of memcache). So, it is imperative for software developers to understand the working and development of this technology.