Cool threads and hot nerves
Dirk Wetter, Dr. Wetter IT-Consulting
Sun Microsystems – or to be more precise Sun's CEO, Jonathan Schwartz – invited developers and customers to test their systems under Ubuntu and write a report about their experiences.
Because the potential reward for the feedback was tempting and a part of my business is do system consulting/research and wrap it up in technical reports I thought "this seems to fit". There was the choice between Sun's Galaxy (AMD64 CPUs) and the Niagara systems (T1). So I ordered a T1 system through Sun's Try & Buy program a great offer also for customers who are unsure whether the system they have in mind is the right one.
Basically the first Linux architectures I used more than 10 years ago for private purposes were not Intel-based (m68k and alpha). However I felt always that the non-Intel architectures were technically not as good as the main i386-line. Out of curiosity I wanted to find out some years back how good the Linux/SPARC port is. IIRC the first try 1997 on an SPARC 10 was the best shot, I tried one year later on an Ultra 5, around 2001 on an E4000 and an E6500 with 28 CPUs. "There was no point" was always my conclusion, "it doesn't buy me anything. Solaris runs great on Sun's SPARC platform, whereas Linux does not." Moreover on Solaris/SPARC in addition to free software there's commercial software from ISVs which will never exist on Linux/SPARC.
June 2006, five years later, Solaris 10 Update 2 hit the road which added an innovative filesystem (ZFS) to Solaris 10 which was in my experience already a good step forward in development. The basis – OpenSolaris, open sourced under CDDL, had in the same month it's first official birthday. Coincidentally also in June 2006, Ubuntu Linux Dapper Drake was released. As the announcement mentions, one of the supported platforms being Linux/SPARC. The UltraSPARC T1 architecture was even certified, whatever that means.
In my five years of Linux/SPARC abstinence I assumed a lot happened and despite the fact that Solaris 10 was there, I thought to give it another try. This is where I started from.
The plot I outlined was the following: I wanted to run a performance comparison, mainly disk and network I/O, and a variety of application benchmarks, some locally like big compiler jobs and others like LDAP and WWW over the network. The competitors were supposed to be Solaris 10 HW 6/06 on one side and Ubuntu 6.06 LTS on the other side. In addition to that, the plan was to spend some words what from the system administration perspective each operating system is doing good and what could be done better. During my career I spent enough time with both Linux and Solaris in data centers, so this would be the cherry on top.
The test environment is basically as in all other cases e.g. for
enterprise tests: There's a switched Gigabit network, a Gigabit
router, a DHCP server handing out
fixed IP addresses and network settings, a central syslog server. For
this test worth to mention that in the lab is a ready-to-go jumpstart
environment: a jumpstart server – an old
but for this purpose still useful Ultra 10 – which is replying to
RARP requests, a boot parameter (rpc.bootparamd), TFTP and NFS server, so that
classical installations with just boot net work. In addition
the boot net:dhcp-method is handled by the DHCP server above
who is also serving boot images via TFTP. An odd thing which is worth
to mention upfront: The DHCP/TFTP server machine runs Suse Linux which
comes with tftp-hpa ("hpa" are the initials of the author). This
TFTP server doesn't work with any Sun OBP I have tested so far if you
use boot net. It just loads
the image and bails out which others experienced, too. The Linux-based TFTP server
is in this case only responsible for providing images to Sun machines
started with boot net:dhcp.
There was also the idea to compare Linux and Solaris on Sun's
Galaxy platform. I decided however that this would have been less
fun: On AMD64 systems Linux runs well, Solaris x64 I would
very much suppose, too. I like challenges and since this appeared
to me too much of an easy goer my choice fell on the UltraSPARC
and 4 USB 1.1 ports, 2 in the front, 2 in the back (why not USB 2.0?). Overall the impression was/is: a niece piece of hardware! Apart from the data center features mentioned above if you have to swap parts: Service informations are on the top cover, you can easily remove it without a screw driver. The same applies for swapping one of the three fans in the front. And no loose cables anywhere. Not only that service is made easy, it's also just great to look at it. On the inner side of the top cover there are even instructions of a more tougher nature: how to remove the motherboard, the SAS backplane and several other boards.
In the the lab everything was hooked up accordingly – the
serial console to a console server, one of the PSUs to a baytech powerstrip, the first
Ethernet port (NET0) to the GE switch, the Ethernet console to a switch in the management network.