Symmetric multiprocessing architecture (SMP) is considered to be a major substitute for commercial servers within a scope of certain scale for many upcoming years to come. The question arises here is how to construct scalable systems, as they needs to be reinitiated. A major limiting factor for SMP is not the capability; instead it is complex to construct an SMP with massive central processing units disbursed in large number of physical casings. Again a question arises here again i.e. how to substitute from SMP for a scalable corporate solution that will facilitate organizations to achieve competitive advantage. Likewise, the SMP technology will be in the market till its technological and economic advantages that can be implemented on more than one SMP. These SMP’s are called as Multiple SMP’s (MSMP). The MSMP is dissimilar from other scalable architectures/systems. Likewise, MSMP is integrated with one or more SMP’s that is different from other scalable enterprise solutions. If we assume a WildFire system overview in relation to the MSMP, there are in total 16 dual core processors and I/O boards. These I/O boards are exchangeable so that he system equipped with one I/O board can run 30 processors i.e. 15 dual central processing boards. The GigaPlane is compatible to support 50 million transactions per second with 112 outstanding transactions with a peak bandwidth rate of 3.2 gigabytes per second. Moreover, latency is calculated to 252 nanoseconds for the total shared memory. The WildFire is compatible to support full cache coherence along with Total Store Order (TSO) that is similar to other Sun based architecture. (Mauro & MacDougall, 2001) The default option of WildFire is the support of “Cache-Coherent Non-Uniform Memory Access” (cc-NUMA) is constructed from massive nodes. For the most layers of the operating system, Wildfire is recognized as a single system, as I/O and DMA are managed by the WFI. Moreover, the allocation of memory is segregated by nodes, wherever required, local memory is utilized for fulfilling processes of memory allocation (Mauro & MacDougall, 2001).
Memory shared throughout the system is reinforced by multithreading and individual processes along with explicit memory sharing in between processes (Alrahahleh & Owaied, 2012). Wherever there is a possibility, threads associated with multithreading processes are placed together within the same node. (Mauro & MacDougall, 2001) In case of more threads in a process, processors start to extent multiple nodes that may facilitate sharing of process memory across the WildFire interconnect. For minimizing and latency and the traffic associated with remote-memory, (Alrahahleh & Owaied, 2012) Coherent memory Replication (CMR) allocates a unique location for every address. Precisely, CMR grants permission for an SMP to provide local shadowing. (Alrahahleh & Owaied, 2012) Likewise, the operating system extracts pages for duplicated data and coherence is managed by the hardware layer located at 64 byte block. For minimizing or eliminating performance issues associated with memory load, WildFire can switch itself with cc-NUMA and CMR on page-by-page and node-by node basis. Likewise, all the newly created pages are established as NUMA. The operating system utilizes the integrated hardware that counters for determining the page that needs to be switched from CMR to cc-NUMA (Mauro & MacDougall, 2001).
For discussing the MSMP architecture, we have used different operating systems in terms of simple latency comparison (RAJARAMAN & RADHAKRISHNAN, 2007). We will derive an MSMP approach of Wildfire and compare it with other operating systems. We will follow a DSM methodology and compare it with Origin 2000 developed from SGI and NUMA Q developed by Sequent. These three systems demonstrate unique approaches for constructing DSM systems. Every DSM node is incorporated with two R 1000 central processing units that are integrated with memory controller. Likewise, the state of the directory is located within the data residing in the DRM banks. This will allow the massive directory state to be accessible easily, in case of no data in the primary node. Although, the directory lookup within the DRAM will be located on the critical path in order to access the garbage data available in the cache in other node. Likewise, every central processing unit holds a big cache size; however, there is no cache node that is also known as remote access cache to facilitate a location of an extra node. For managing the massive traffic coming from the Internet, high bandwidth connection is deployed within the distributed routers. Therefore, the system bandwidth equipped with 32 central processing units incorporates 6.4 gigabits per second (RAJARAMAN & RADHAKRISHNAN, 2007).
Whereas, NUMA-Q is constructed from little branded blocks equipped with four P6 processors residing in each node. Every node is equipped with a cache of 32 MB that is shared among all central processing units. As mentioned earlier that the nodes are constructed by SMP nodes bundled with its own memory controller, directory do not resides with the data. The local latency of WildFire is smaller as compared to Origin and consequently, larger than NUMA-Q. Likewise, for traffic coming from remote sites, the sequence is opposite by placing Origin on the first place with WildFire and NUMA-q to follow. Moreover, the latency in remote connections for Origin and NUMA-Q are based on the size of the system, as compare to WildFire. Likewise, Windfire’s enormous node provides a low cost solution. The system that is constructed for handling massive nodes will not experience latency in remote connections unless the system size becomes larger than the node size. A generic SMP application is not optimal for non-uniform architecture with memory access time. Moreover, an advantage that WildFire shares is the reduction of random cache-to cache delay in large nodes. Likewise, this enables the large latency for remote nodes and empowers them to be the fastest data migration solution.
We have discussed SMP and MSMP in the context of Wildfire developed by Sun Microsystems. We have concluded that WildFire is the most suitable solution for sustainable and feasible architecture for handling workloads. Moreover, it also handles multi-tasking, load balancing and minimizing the overburdening of memory within the database. Furthermore, we have also compared and contrasted three products for SMP and MSMP i.e. Origin, WildFire and NUMA architecture
Alrahahleh, A. M., & Owaied, H. H. (2012). Scheduling model for symmetric multiprocessing architecture based on process behavior. International Journal of Computer Science Issues (IJCSI), 9(4), 77-84.
Mauro, J., & MacDougall, R. (2001). Solaris internals: Core kernel architecture Sun Microsystems, Incorporated.
RAJARAMAN, V., & RADHAKRISHNAN, T. (2007). Computer organization and architecture PHI Learning.