Linked by Thom Holwerda on Wed 14th Jun 2006 16:05 UTC, submitted by _DoubleThink_
Talk, Rumors, X Versus Y This paper tries to compare Linux vs. Solaris. Its author comes to many conclusions, among which this is one of the more interesting: "All-in-all Solaris is a powerful, stable, conformant-to-standards OS that can run many open source applications as well as Linux, and some (mainly multithreaded applications) better than Linux. Like in the cases of Red Hat and Suse, the cost of support is extra, but it is more reasonably priced. Security patches are free which makes Solaris similar to Windows."
Thread beginning with comment 133529
To view parent comment, click here.
To read all comments associated with this story, please click here.
RE[3]: Solaris scales better
by ormandj on Wed 14th Jun 2006 21:26 UTC in reply to "RE[2]: Solaris scales better"
ormandj
Member since:
2005-10-09

People no longer seem capable of differentiating between the two. Most people here wouldn't know what NUMA was if it bit them in the tail.

Horizontal scalability is all people that have grown up in the late 90s/early 00s know. Vertical scalability is a thing of a past to them (actually, most of them have seem to have no idea what it is..)

So, to them, they likely aren't comparing apples and oranges. They just don't know the difference between apples and oranges, and haven't had a taste of either, so they just talk out their tail-end about what they *think* based on their obviously highly-educated guesses.

Reply Parent Score: 3

RE[4]: Solaris scales better
by kaiwai on Thu 15th Jun 2006 01:17 in reply to "RE[3]: Solaris scales better"
kaiwai Member since:
2005-07-06

So, to them, they likely aren't comparing apples and oranges. They just don't know the difference between apples and oranges, and haven't had a taste of either, so they just talk out their tail-end about what they *think* based on their obviously highly-educated guesses.

Babe, its known as what I like to call, 'the google effect' where by they scream, "hey! google can do that!" when ever they wish to diss Solaris, big iron midframes/mainframes.

Bascule had a good post explaining the differences, comparing the Google approach to a big iron running Oracle - too bad the Linux fanboys (and girls) can't be bothered getting tuned into it.

Reply Parent Score: 1

RE[5]: Solaris scales better
by ormandj on Thu 15th Jun 2006 01:58 in reply to "RE[4]: Solaris scales better"
ormandj Member since:
2005-10-09

"Babe"? Please, out of respect for me, don't call me that.

Other than that, yes - concur, that's generally what happens. Latency doesn't mean a thing for some parallel jobs, to others it does. The jobs cited for examples don't require extremely low latency between processors.

Something that requires in-order parallel processing wouldn't be well suited to the "Google" style of clustering cheap boxes. When a reply might take an extra 1ms to get where it's going, it can mean the difference between working and not working. Oh well. There are things somewhere in the middle (SGI), and then there are highly vertical scalable monster boxes from Sun. Of course you can find examples of each from other manufacturers, just using these as well known examples.

Reply Parent Score: 1

RE[4]: Solaris scales better
by CrLf on Thu 15th Jun 2006 03:17 in reply to "RE[3]: Solaris scales better"
CrLf Member since:
2006-01-03

"People no longer seem capable of differentiating between the two. Most people here wouldn't know what NUMA was if it bit them in the tail."

This pops up every single time one talks about Linux scalability to large number of CPUs, and it's annoying...

SMP and ccNUMA ("cc" - cache coherent - being the important thing here) are basically the same thing, a bunch of CPUs running a *single* system image, which means the same locking contention problems appear on both architectures, albeit with differences.

The difference between the two is the memory access, which (like the name says) doesn't have uniform latencies/speeds for every address. ccNUMA schedulers are basically SMP schedulers that have to take memory locality into consideration.

SMP and ccNUMA are hardware architecture choices: on the SMP side the designer wants to make things simpler for the OS and optimizes for processes accessing mostly the same data, on the ccNUMA side the designer wants to maximize memory access bandwidth and optimizes for processes accessing their own data.

But in any case, all processes will have to make system calls, do IPC, and that where kernel scalability comes into play (together with making the best choices for memory allocations, CPU affinity, etc).

As to actual hardware...

A 512-CPU SGI Altix physically is a bunch of boxes with a couple of CPUs and memory, connected by an high-speed bus (NUMAflex on the Altix, if I recall correctly), but a big classic SMP box is also a bunch of cards with CPUs connected by an high-speed bus (but without any local memory).

The Opteron is ccNUMA, the Xeon isn't. What's the difference between scaling well to an n-way Xeon machine or an n-way Opteron? Get the picture?...

Moral of the story: 512-CPU SGI beasts AREN'T CLUSTERS.

http://oss.sgi.com/projects/numa/Linux_Scalability_for_Large_NUMA_O...

The big problem here is Solaris fanboys are so used to their "nothing scales like Solaris" idea that they mix hardware issues with software issues.

Reply Parent Score: 2