Wandering Thoughts archives

Thesis: SMP is a failure for most Internet servers

I used to be a big proponent of getting SMP PClone servers instead of uniprocessor ones. Nowadays, though, I'm not so sure, and I now actively tilt towards UP machines.

The advantage of SMP over UP is more CPU capacity and to some extent larger caches (if the OS scheduling works out right). But Internet-facing servers are rarely hurting for more CPU; often you really want more IO and maybe more memory.

(This assumes that the server is not grotesquely overconfigured to start with, which happens more than I'd like to think, especially because vendors rarely sell low-powered servers.)

And the downside of another CPU these days is not just in the extra cost. The CPU is one of the largest sources of heat (and consumer of power) in modern servers, which is now an issue of serious concern in many data centers.

The one wart I can see in this is widespread use of SSL and other cryptography, since that can eat a bunch of CPU. On the other hand, experimentation suggests that a two year old uniprocessor can do straightforward bulk transfers at 100 mbits/sec while using less than half the CPU. (Admittedly, bulk transfers are only half the story, as I believe that doing the SSL handshake eats more CPU than a stream cypher.)

A theoretical advantage of SMP is that the OS can drop currently unneeded CPUs into low power modes or turn them off entirely. But for this to really work out you need the CPUs you're using to draw less power than a normal UP machine's CPU, so that you're saving over just having a plain UP machine to start with, and my impression is that no one sells PC machines with low power consuming SMP CPUs.

(In general SMP capable CPUs always seem to be at the high end of Intel and AMD's product ranges, with the power consumption that goes with that.)

The problem of IT winning arguments

Yesterday I wrote about how sysadmins are overhead. One of the consequences of this is that IT does not actually have real power. The real power accrues to the people who are doing the organization's work, whereas IT is just keeping the machines dust-free.

Remembering this will make life less explosive, because the people with actual power can usually squash IT flat in thirty seconds if push comes to shove. Generally you don't want it to go that far.

(This is part of the core of the problem with punishing people at universities. Tenured professors are not only tenured, they're what the university is about.)

Where IT manages to successfully call the shots over the objections of people actually doing stuff, it does so through what is ultimately fear, usually fear of security problems or fear of expense. The problem with this is that people resent being scared (or equivalently, held to ransom), even if you are correct.

And this assumes that you've actually scared the people. Often, IT has only scared management, not the actual people that their policies affect. (This is always the case if IT's argument is 'the other way costs too much money', since that's a management concern.)

(The other problem is that people don't stay scared, so IT has to try to re-scare them every so often.)

The best way for IT to sell things is to find some way that they make life better and more convenient for people. 'Do this because there is no viable alternative' is a poor second.

Sidebar: the problem of money scares

Fear of spending money is an especially pernicious approach for another reason; it's often very difficult to quantify the benefits of spending the extra money, and thus difficult to defend doing so. It's easy to produce figures for how much management could save by locking down worker desktops (you just show how much time support staff would save and how much that time nominally costs), but much harder to put a clear dollar figure on not doing so, and thus you wind up with the virtual furniture police.

Another annoying RSS feed trick

Following up my previous entry on this subject, I've run across another clever way to do partial feeds: include only the first paragraph of every entry without any indication that there's more. This works especially well if your first paragraphs are pretty much self-contained, so that they're easy to mistake for full entries.

If people haven't browsed your blog directly (perhaps because they're reading you through an aggregator), you may be able to fool them for some time. Won't they feel silly when they later realize how much they've been missing?

For bonus points, have your entries cover a variety of subjects (in separate paragraphs or sections), so people can't just read the first paragraph in your feed to get a sense of whether they want to click through to read the rest.

(In my particular case I only found out about this when I clicked through to see if an interesting looking entry had gotten any comments.)

In general, partial feeds still puzzle me for old reasons: fundamentally they're a disincentive for would-be readers to actually read your material. Some people will simply not subscribe to your syndication feeds; others will subscribe but not click through all the time. Net result: less readership. This strikes me as an odd thing to aim for.

If you are worried about overwhelming people with a stream of vastly different entries, most of which they won't be interested in, there is a simple solution: offer subset feeds. WanderingThoughts does, and I know there's at least one person who's taken advantage of it.