Softpanorama (slightly skeptical) Open Source Software Educational Society

May the source be with you, but remember the KISS principle ;-) Softpanorama Search

Solaris vs. Linux: Framework for the Comparison

by Dr Nikolai Bezroukov

3. Four major areas of linux and Solaris deployment

Generally we need to distinguish between several major areas of linux and Solaris applicability:

1. Development workstation. Here linux dominates (due to absence of virtualization capabilities it is often used under VMware and now with newer CPUs under Xen3) because this is its area of specialization and Solaris on UltraSparc can be competitive only in configurations with SunPCi cards and Solaris on Intel only under VMware or on newer CPUs Xen. The ability to run Windows is an essential requirement for most development workstations and with the current dual-core CPUs there is no reason not to run two OSes on the same desktop or even laptop (which due to Vista requirements is now often configured with 2G of memory). That's were VMware was extremely successful and it is actually sad that Sun brass failed to buy the comply and it was snapped by EMS, the company with which it have rather little synergy. Workstations are the area were virtualization is becoming standard tool of the trade in a large enterprise environment. We will discuss it in more details below.
    
2. Low end servers: single or dual core CPU servers with 4G or less of RAM and with the total cost less then $6K (servers with the cost less then $5K often are considered to expense in large enterprize environment, so the price of $5K+ is the minimum under which you can capitalize and depreciate them).  From the technical standpoint Linux has advantages on the low end as this is its main base. Sun competes mainly on the high end of this price range with T1000 and X4100/X4200, but suffers from too rigid component selection (too few choices among installed hard drives is a real problem for Sun). Solaris kernel is biased toward multiprocessor machines and until recently Solaris performance suffered on a uniprocessor and dual processor servers. With improvements introduced by Sun in Solaris 10 the jury is out and on multicore Opteron servers it might be more competitive than before.  One significant drawback of Solaris in this area is a very weak integration with Microsoft server infrastructure.  Novell Suse beats Solaris in this area and that fact along might spell troubles for the future of Solaris on low end despite its real or imaginable advantages. Like Talleyrand used to say about such strategic errors: "It is worse the a crime this is a blunder".
    
3. Midrange servers (2-8 CPU servers often with multiple cores per CPU, and 8-64G of RAM with total cost from $6K to $100K). Solaris on UltraSparc used to have traditionally strong presence in this segment and was the development platform for 64-bit Oracle (home field advantage), but with the release of the newest line of Intel hardware (Duo and Quattro CPUs with 1.33GHz memory) linux beats Solaris 10 on UltraSparc by a considerable margin (sometimes to the extent that two UltraSparc servers can be replaced by one similarly configured Intel server).  The fact that Oracle now promotes its own distribution also tip the scales in favor on linux.  Still Solaris is definitely more stable on high loads and works better then Linux with larger memories (starting probably with 64G of RAM) and with large number of CPUs (over 8).  But if we consider Opteron-based servers only, Solaris X86 is competitive with linux and might even be superior in such metric an stability under high load. Actually in case of Red Hat you need a different, twice more expensive, distribution (RHEL AS) for such configurations but this does not matter much for this class of deployment as everything is pretty expensive here.
    
4. High end servers (more then 8 CPUs and more then 64G of RAM with cost starting from over $100K). This sector is traditionally called "mainframes" although in a narrow meaning mainframes are descendants of famous IBM/360 series that for twenty years (1960x-1980x) dominated the computer landscape. Currently linux has no significant presence in this segment of the market. The Linux kernel is much less tested with those configurations and as such is attractive neither for large enterprises nor for major vendors like IBM and HP (both of them have their own flavors of Unix for this segment, which work much better then linux).   Solaris (along with AIX and HP-UX) is the king of the hill in this deployment area although with Windows 2003 Microsoft got a foot in the door too. 

There is also new fashionable area called  virtualization that falls in between those classes. It is essentially and attempt to imitate functionality of multiple low end servers on a mid-range or high end server.  It is connected with the trend to consolidation of  servers with extremely low load (circulating air in large corporation IT jargon) that is in a dialectical way signifies the return of good old mainframes on a new level. Actually IBM mainframes survived all adversities and now, more then 40 years after their launch, can proudly host Linux servers. This shrewd IBM move helped to save classic VM/CMS operating system which naturally became VM/Linux.

Currently AIX and VMware are the most popular choice for virtualization based server consolidation but recently Xen made inroads into VMware turf (although the development was financed by Microsoft, it is now supported by IBM).

 Generally comparing servers the following two factors should be considered for all classes of servers:

• Home field advantage
• General conservatism of the large enterprise environment

Home field advantage

Modern applications are extremely complex with huge code bases. In a sense they are modern engineering marvels. That means that other things equal the key to right selection of the platform is the selection of hardware platform and OS that is used for development of a particular application. If application is developed on a particular platforms and OS then this OS has an important advantage that we will call "home field advantage".

That means that for Java applications Solaris should be considered the No.1 platform as the quality of implementation of Java on Solaris is the highest due to the fact that it is the development platform for Sun developers.  For Java applications T1/T2 based servers provide an unmatched performance beating much more expensive systems including IBM P5 550 4 core 1.9GHz Power 5+ in SPECjjb2005 benchmark.  It has interesting implications that were first noticed by Paul Murphy in his paper "The megahertz myth and the UltraSPARC T1"[Merphy2006]: 

For the same reason for most open source application linux is the preferable platform as it serves as a development platform.

Another relatively important factor is the level of interest of the vendor in open source applications. Actually open source applications are extremely important for Sun as the best of their low end CPUs (T1 chips) are multicore and usage of open source application allow completely bypass the issue of paying for multiple cores. This is not true for many commercial applications. As such chips are becoming more and more mainstream they might help to drive the level of support of open source applications and scripting languages in Sun.  Sun now support more then a dozen open source applications installed by default ( "out-of-the box" ) on Solaris with the most recent addition of Postgress [Sun2006a]:

Fully integrated into Solaris 10 with flexible support offerings from Sun, Postgress on Solaris 10 is the open source enterprise database platform of choice. 

For all "tier-1" open source applications the level of  support is equal to the level of support its Sun's own software. And that is a very positive development.

General conservatism of the large enterprise environment

The fact that it in large enterprise environment linux is used only for rather narrow set of roles with web server farms as the most prominent. Beyond this role it its growth is very slow. In many cases RISC-based servers (not necessary Sun) are so entrenched that to replace them with Linux of Intel (or Solaris on Opteron) is very costly and can be done only during big corporate earthquakes like mergers and acquisitions.

I would like to stress it again  that in large enterprise environment low end UltraSparc based servers with Solaris are competitive both with linux servers and Solaris on Opteron for several reasons unrelated to CPU, relative quality of those two OSes, or even transactional benchmarks. The main reason for such situation is that proliferation of Unix flavors negatively affects large enterprise environment. And usually linux increases the diversity of enterprise environment not by one unit but by two as very seldom old flavors of Unix are completely displaced and there is a real  danger that both Suse and Red Hat will become eventually deployed. 

All naive or crooked stories about how corporation ABS saved a couple of million dollars by deploying linux are what it is: naive or crooked. In reality introduction of linux is often dictated by fashion and/or by the inability of higher management to distinguish between real cost cutting measures and fake one: the cost of adding another flavor of Unix in a large enterprise mix easily nullifies any savings in hardware unless one or two existing Unix flavors (for example HP-UX and AIX or HP-UX and Solaris) are completely eliminated from the mix.  In this particular area large enterprise environment is quite different from Internet startups; unlike startups large enterprises has history of Unix deployments and significant (and often expensive) installed base of classic enterprise flavors of Unix (AIX, HP-UX and Solaris; often all three of those).

As I noted before this problem in less severe form exists for Solaris too: Solaris SPARC and Solaris on Opteron is still two different OSes (not then much as Red Hat and Suse, but still there are differences because hardware platform are so different). For example Open PROM is not available on Opteron. So objectively there is a cost of introduction additional flavor of Solaris on Opteron into Unix ecosystem too. It is just less then in case of linux.  In fact my impression is that the more administrator is qualified in Solaris on UltraSparc the more painful are those small differences that exist between Solaris on UltraSparc and Solaris on Opteron.

That means that large corporate environment generally favors Solaris on SPARC to Solaris on Opteron and paradoxically many Sun Opteron server are used with either Windows or linux as both those OSes have a tattoo on their foreheads: "made for X86 platform". 

Of course this is a conservative position and it does not take into account price/performance differences between UltraSparc and Opteron that is another driver of change.  But if  integer or floating point performance is not crucial for a particular application (which is true for all I/O bound applications) and benchmarks are close there is very little reason to jump the UltraSparc boat: there is nothing in Intel X86 architecture that is intrinsically attractive other then the combination of low cost and high performance bought by hundred of billions spend by AMD and Intel on keeping alive the technology which belongs to late 70th of the last century. I hope that in 2007 Sun eventually will add  new  CPUs to its lineup as a result of its join product development agreement with Fujitsu.  With the Fujitsu PrimePower UltraSparc systems outperforming Sun's SunFire systems (PRIMEPOWER450 (1870MHz)  has SPECint2000=1344) and matching Opteron systems this merger of technologies sounds very reasonable and timely.