Top Links You Must Click On
Open Source PaaS for Parallel Cloud Application Development
Architectural advantages of cross-layer optimized parallel computing PaaS
By: Mark Sandstrom
Feb. 18, 2013 08:00 AM
Improving application program performance will require parallelizing the program execution at ever finer granularity now that the processor clock rates are no longer increasing. However, even in a per-application dedicated computing environment, the parallelization overhead is known to place a limit on how much application on-time throughput performance increase can be achieved via higher levels of parallel processing. The throughput-limiting impact of parallelization overhead will be significantly amplified when executing multiple internally parallelized applications on dynamically shared cloud computing environment, since the allocation of processing resources to instances and tasks of any given application cannot be done in isolation, but instead it needs to be done collectively across all the applications dynamically sharing the given pool of computing resources. There thus is an urgent need to solve this complex challenge of developing internally parallelized programs for dynamic execution on shared cloud computing infrastructure, if we expect to be able scale the performance and capacity of cloud hosted applications going forward.
PaaS Approach for Parallel Cloud Computing Challenges
This new platform architecture will involve a parallel program development environment for producing application executables based on actors that can be efficiently mapped for concurrent execution on processing cores of dynamically shared manycore arrays. In addition, the parallel cloud computing platform will need an execution environment that, besides executing the program instructions on processing cores, takes care of the dynamic parallel execution routines on behalf of the applications, so that the processing cores are used for executing the actual client programs instead of system functions. If the hardware of the manycore processor fabric did not handle the parallel execution routines, eventually the system would be just managing itself rather than providing increased user application throughput as the numbers of processing cores and applications and tasks sharing them are scaled up. This difference in enabling application on-time throughput scalability is illustrated at Figures 4 and 5 (in Ch. 3 Architectural Advantages of Cross-layer Optimized Parallel Computing PaaS).
The Platform Architecture
Figure 1: Overview of the parallel cloud computing PaaS
As illustrated in Figure 1, the open parallel computing PaaS has an open-sourced parallel program development environment and a dynamic parallel execution environment. The development environment allows the platform users to produce executables of their application programs that are made of segments (tasks/actors/threads) that can execute concurrently on parallel (incl. pipelined) processing cores. The execution environment provides, besides an array of processing cores for parallel execution of the user program tasks, the capabilities to dynamically map the highest-priority ready application task instances for execution on their assigned cores. While the state (e.g. which application task instance is mapped to any given core of the shared resource pool at any given time) of the execution environment is highly dynamic during the application runtime, in the platform architecture per Figure 1, the execution environment provides for the application programs (and their developers) a virtual static view of it; any given application can assume that each of its task instances is always active and mapped to a virtual static core in an array that is virtually dedicated to that single application. The hardware automated parallel processing runtime routines of the execution environment per Figure 1 hide (to the desired level) the dynamic details of the processing hardware from the applications as well as the development environment software, thereby providing a higher level, simplified abstraction of the execution environment for the software. The raised level of the software-hardware interface per Figure 1 enables greater productivity for both realizing the much needed open, comprehensive parallel programming environment as well as for developing parallelized applications for cloud deployment, as software does not need to be concerned of the dynamic parallel execution details.
The IDE software can be hosted with the same parallel computing PaaS as the applications it produces. This practice avoids the need for cross-compiling, and enables straightforward and rapid-cycle interactive testing, debugging and optimization of the parallelized application programs, as well as deployment and scaling releases of the user applications.
The software of the parallel cloud computing PaaS is developed and made fully available as an open-source project, and can be integrated with popular open-source IDEs. In essence, this software project is to add the parallel programming development tools to the major open source IDE and PaaS code bases, while utilizing (and further developing) the applicable existing features from them.
The openness of the promoted comprehensive parallel cloud computing PaaS, besides the open-source IDE, is also manifested via the simple, open standard interface between the development and execution environments of the PaaS architecture. This open standard interface enables any user to host the IDE as well as the parallel program executables produced by it anywhere, e.g. at the user premises, or with ThroughPuter or any 3rd party. Alternative I/PaaS providers furthermore are encouraged to support the execution/development environment interface via their respective implementations of either side or both sides of that interface. The customizable, open-source IDE and open-standard interface to the execution environment provides the users and collaborators a flexible and productive way to approach the major, must-solve parallel cloud application development and execution challenge that is facing much of the software industry and its customers.
The efficient dynamic parallel processing features - which will be critically needed as user application throughput requirements begin exceeding what is available from conventional sequential execution models, and as the parallelized applications will be cloud hosted -- of this open-source software based PaaS are delivered by an execution environment that provides the necessary, dynamic parallel execution core to application task instance allocation, task instance to core assignment, and inter-task communication capabilities. These critical parallel execution capabilities are an integrated feature in the ThroughPuter hosted commercial PaaS offering.
For reference on ThroughPuter's implementation of the dynamic parallel execution environment for the open parallel computing PaaS architecture per Figure 1, the below Figure 2 shows ThroughPuter's realtime application load adaptive manycore processor architecture, highlighting its hardware logic automated operating system functionality enabling a number of customer application programs to securely, dynamically and cost-efficiently share the processing capacity of the manycore processor hardware.
Figure 2: Reference diagram for ThroughPuter manycore processors with hardware-automated multi-user parallel processing optimized operating system.
ThroughPuter's hardware operating system, manycore fabric memory and I/O subsystems are largely responsible for the architectural security, performance and cost-efficiency benefits of the ThroughPuter PaaS. However, the IDE hides the actual execution environment features from the user; the user does not need to be aware of the novel hardware-implemented capabilities of this dynamic parallel execution environment in order to realize the performance benefits.
A possible core to application task/instance allocation scenario over a few core allocation periods (CAPs) is illustrated in Figure 3 below, along with associated highlights of the feature benefits.
Figure 3: Dynamic core to application task instance assignment scenario in the execution environment of the PaaS per Figure 1.
In reference to the dynamic core assignment scenario per Figure 3, it shall be remembered that in the PaaS architecture per Figure 1, the software does not need to be aware of the dynamic parallel execution matters, but can instead maintain a virtual static view of the execution environment where each possible application task instance is constantly mapped to its virtual dedicated core, thus simplifying both the development environment as well as the application software while improving the development productivity and runtime performance.
Architectural Advantages of Cross-layer Optimized Parallel Computing PaaS
Figure 4: Scalability problem in parallel cloud computing due to the system software overhead.
As illustrated in Figure 4, when relying on system software for managing the parallel execution, there is a limit to scalability of cloud computing platforms in the parallel processing era, as application performance improvement begins requiring ever finer grades of intra-application parallelism (i.e. more tasks and task instances per application). This is due to that the parallel processing system software overhead (the need to dynamically coordinate and manage concurrent tasks and parallel processing resources) increases with the number of applications as well as their tasks and instances, and the number of processing cores being dynamically shared among them. When relying on system software to handle parallel execution routines, the processing capacity of a given manycore processor (array) is split between processing user applications and system software, with the rate of system software increasing with scale, at the expense of the user applications. This causes that after some point, the system-wide application processing on-time throughput (the product of number of cores and the percentage of the cores' processing capacity available for user applications) will begin decreasing as the system is scaled up (by adding processing cores and parallelized applications and their tasks sharing the cores).
To solve this fundamental challenge affecting the scalability of cloud computing in the parallel processing era, the processing hardware needs to raise up to this challenge and handle the parallel processing routines in the hardware of the manycore processors, so that the processing cores will be optimally used for processing the user applications (rather than for processing the system functions, and/or be locked to low utilization due to non-load-adaptive allocation).
The impact of the hardware automation of the parallel processing system functions in the hardware of manycore processors per Figures 1-3 is illustrated in Figure 5 below.
Figure 5: Scalability solution for parallel cloud computing delivered via automating the parallel execution system functions in hardware.
The enabling of scalability of cloud computing platforms and cloud applications' on-time processing throughput in the emerging era of (inter and intra) application parallel processing by the execution environment model per Figure 1 serves as a further compelling reason for concentrating the efforts to address the popular parallel programming challenge via the herein presented open parallel cloud computing platform model. Parties interested in collaboration to realize this much needed comprehensive, open source parallel cloud computing platform can contact ThroughPuter via firstname.lastname@example.org.
Reader Feedback: Page 1 of 1
Enterprise Open Source Magazine Latest Stories . . .
Subscribe to the World's Most Powerful Newsletters
Subscribe to Our Rss Feeds & Get Your SYS-CON News Live!
SYS-CON Featured Whitepapers
Most Read This Week