Update Makes Red Hat Enterprise Linux 4 More Scalable

Published: March 21, 2006

by Timothy Prickett Morgan

Early on, the scalability of the Linux operating system--or more precisely, the lack thereof--was a hindrance to the adoption of Linux as a strategic platform for enterprise-class applications that needed more than one or two processors. But over the years, the scalability of Linux has been stretched such that it can span much of the very scalable hardware to which it has been ported. But, there's always room for more, and that is why Red Hat last week put out an update to its Enterprise Linux 4 to extend SMP scalability even further.

With Linux 2.2, the kernel could scale really well on two-way symmetric multiprocessing (SMP) configurations, which have a shared cache memory architecture that makes two processors share the same memory space and therefore allows them to run one copy of the operating system. SMP architectures on big RISC/Unix boxes have been in the dozens to hundreds of processors for the better part of six years now, and Linux has been a laggard. But it is catching up. With Linux 2.2, SMP scalability was technically to four-way SMP, with two-way SMP working very efficiently. With the Linux 2.4 kernel, up to eight-way SMP was technically supported, and four-way SMP worked reasonably well. With Linux 2.6, depending on the implementation and the server architecture, either 16 or 32 processors in an SMP or NUMA cluster were supported, and eight-way SMP functionality was regarded as being efficient and generally as good as what Unix or Windows could deliver.

But multicore architectures are going mainstream now, and SMP and NUMA clustering are the only ways left to boost system performance now that we have hit the thermal windows on processor designs and cranking the clock speed is no longer an option. So any operating system worth its salt has to use the SMP scalability inherent in these multicore chips and also span the larger SMP and NUMA servers they allow when they are glued together into larger systems by advanced chipsets.

To that end, RHEL 4 Update 3 has big time scalability improvements for this implementation of Linux. The operating system includes a new "largesmp" Linux kernel specifically designed for 64-bit Xeon platforms from Intel and compatible 64-bit Opteron platforms from AMD. This largesmp kernel, which you exchange for the regular kernel, can support up to 64 logical processors on these boxes. That can mean a 32-socket machine with dual-core processors, but because Intel supports HyperThreading simultaneous multithreading in the Xeons and AMD does not in the Opterons, ironically that means the Xeon core count in the fully extended SMPs get cut in half compared to the Opterons. HyperThreading makes a single core look like two logical cores to the operating system, and boosts performance by 20 percent or so for that core. But it also means that when Intel moves to quad-core processors in early 2007 for servers, with each core supporting HyperThreading, a single socket will consume eight of those logical processors that Red Hat Linux supports. So the top-end Xeon machine that that time will be a four-socket box. Granted, this machine will have a lot of oomph. But the Opterons will be able to support an eight-socket box that, in theory, will delivery quite a bit more performance running Linux and other operating systems.

Red Hat is also bringing forward the existing "smp" kernel for Itanium processors, which was able to span 64 logical processors with Update 2 for RHEL 4, and updating this kernel to run on IBM's PowerPC and Power architecture servers. IBM's largest Power5-based machines have 64 cores and 128 virtual processors thanks to simultaneous multithreading, which means a single RHEL 4 Linux instance can still only span half of a p5 595 server. However, with Intel readying its dual-core "Montecito" Itanium chips for some time this summer, the Linux kernel will have to scale beyond 64 logical processors. To that end, RHEL 4 Update 3 has a technology preview of a different kernel--presumably due in RHEL 5 at the end of this year--that scales up to 256 logical processors on Itanium servers and up to 128 logical processors on Power servers.

RHEL 4 Update 3 also has support for the dual-core Montecito chips, since this is the first dual-core Itanium to come to market. Update 3 also has a technology preview of the OpenIB open source InfiniBand software stack, which Novell is also weaving into SUSE Linux Enterprise Server 10, which is due in May. There are a number of other I/O, security, and device driver enhancements in Update 3, which you can find out about here.