Red Hat revs real-time MRG Linux to 1.2
Better, stronger, faster
Commercial Linux distributor Red Hat today kicked out the 1.2 release of its Enterprise MRG Linux variant for real-time, messaging, and grid computing.
Enterprise MRG was launched as an idea in December 2007 in the wake of Novell's roll-out of its SUSE Linux Enterprise Real-Time variant of its SLES server Linux distro a month earlier. The initial 1.0 release, which was missing the Project Condor grid components, debuted in June 2008. It did have the real-time Linux kernel, which is important for military systems, financial trading systems, and other industrial controllers where you have to strip down the general purpose Linux kernel and make it provide a more consistent response time (with low latency) on transactions.
The Enterprise MRG stack also weaved in the Advanced Message Queuing Protocol (AMQP), an open source and standard message queuing stack akin to the messaging server brokers such as IBM's WebSphere MQ, Microsoft's Message Queuing Middleware, Tibco's Rendezvous financial transaction messaging, and the Java Messaging Service.
In February, Enterprise MRG 1.1 came out, adding the Condor grid components and also adding native InfiniBand drivers (adopted from the OpenFabrics Alliance OFED) and Remote Direct Memory Access (RDMA) drivers for lower-latency clustering across InfiniBand and Ethernet backbones.
With Enterprise MRG 1.2, the company is shifting to Red Hat Enterprise Linux 5.4 as the code base, which was announced in early September. With RHEL 5.4 as its code base, Enterprise MRG 1.2 can now support the enterprise-grade version of the KVM hypervisor that is part of the operating system and that grid and messaging software can run atop KVM virtual machines if customers want to go all cloudy and virtual.
To prove that this can be done without sacrificing performance, Red Hat grabbed the PerfTest messaging benchmark, which it bundles with Enterprise MRG, and dropped it on two Dell R710 two-socket Xeon 5500 servers. Using 10 Gigabit Ethernet links between a box running the messaging broker and another running the messaging client, the system was able to process over one million messages per second.
Bryan Che, manager of product management for the Enterprise MRG product at Red Hat, says this is significant in two ways. First, the test measures message passing from one API to another with guaranteed delivery; on some tests that messaging software vendors use, they fudge on the guaranteeing of the message to boost performance.
But more importantly, for financial-services companies and other organization with messaging architectures at the heart of their applications, the PerfTest benchmark demonstrated that there was only a 5 per cent penalty in going virtual compared to running the same test on the same iron on bare metal. This is probably acceptable performance degradation if the benefits of virtualization (for system resiliency and lower cost of management) are great enough. (The latency on this test was under 200 microseconds, which is kind of slow compared to other tests Red Hat ran, which are discussed below.)
Che said that other benefits that come to Enterprise MRG by virtue of the move to RHEL 5.4 are support for all the current x64 iron from Intel and Advanced Micro Devices (with performance optimizations geared to specific chips), C library optimizations that improve the performance of real-time and messaging applications, and improved OFED drivers for the InfiniBand protocol supporting RDMA access for peppier messaging.
The performance boost in Enterprise MRG 1.2 for InfiniBand/RDMA messaging is substantial. Depending on the message size, as Red Hat shows in these graphics, the company can do from just around 1.2 million to 1.6 million messages per second with an RDMA latency of under 40 milliseconds on reliably acknowledged message transfers. This, says Che, is nearly a 50 per cent reduction in latency from Enterprise MRG 1.1.
Of course, some of the issues with lowering latencies require the tweaking of BIOSes and other hardware features, so Red Hat has cooked up a new tool called rteval that runs on servers, and figured out where the system-management interrupts that have been configured for general purpose computing are slowing down real-time operating systems like Enterprise MRG. Basically, you run the tool and it points the finger at the hardware, but in a useful way that lets you reconfigure it so transactions and messages can move through the system more predictably as well as quicker.
Finally, Enterprise MRG 1.2 has been updated with the 7.4 release of the Condor grid computing software, which comes out of the project that is managed by the University of Wisconsin. Interestingly, Condor 7.4 (which came out last month) supports the running of both KVM and Xen hypervisors on top of the grid software and the Condor grid scheduler has been integrated with the libvirt virtualization tool created by Red Hat so Condor can stop, start, suspend, and provision virtual machines and plunk them on a grid of interconnected servers.
When I asked Che about how Enterprise MRG was doing in terms of generating shipments and revenues, Red Hat PR piped up and said that the company was in its quiet period ahead of reporting earnings for the third quarter of fiscal 2010 ended in November on December 22. So there's no talking numbers until after then. ®
Sponsored: Benefits from the lessons learned in HPC