More Dreamhost wankery

[giza]

[baachus]$ uptime
08:20:13 up 49 days, 2:08, 1 user, load average: 952.32, 740.37, 397.97

Wow, a load of 952... I think that's a record high. Come on Dreamhost, suck just a little harder so you can get to 1,000! Maybe the box will actually catch on fire at that point...

Comments

[zorinlynx.livejournal.com]

What the heck is the machine doing to get such a high load average?

We sometimes see that around here, but usually it's because some NFS server has gone down and dozens of processes are stuck in IOWAIT...

[giza.livejournal.com]

Given how web and SSH are acting on it, I think "What is the machine NOT doing?" is a better question...

[nius.livejournal.com]

Received any Anthrocon memberships from concerned parties in Nigeria who have a million fursuits to smuggle out of the country?

[greenreaper.livejournal.com]

Coming from Nigeria, I'd be a little worried that we'd be getting fursuits made of real fur . . .

[greenreaper.livejournal.com]

Hmm. 49 days . . . that isn't a Windows 98 box masquerading as *IX?

[taral.livejournal.com]

Welcome to why including processes blocked on I/O in load average is dumb.

[giza.livejournal.com]

Actually, I'd say it including it in the load average quite clever, given how poorly HTTP and SSH were performing at that point in time. :-/

[taral.livejournal.com]

From a system diagnostic point-of-view, load average is a near-useless metric because it doesn't tell you much. CPU-bound? Network-bound? Disk-bound? System failure? Resouce limit? What's wrong? I've seen systems where a load average of several hundred is *perfectly normal*.

Linux, unfortunately, has a real dearth of easily accessible performance metrics. :(

[giza.livejournal.com]

> I've seen systems where a load average of several hundred is *perfectly
> normal*.

Oh, can you elaborate? It's not relevant to the current discussion, but now I'm curious!

[taral.livejournal.com]

There was at least one customer that parallelized his I/O by launching thousands of processes and letting I/O pacing take care of the load. The result was a mid-level CPU load and a load average around 400-500.

[giza.livejournal.com]

OMG.

Did you beat him over the head with a book about async I/O? :-)

[taral.livejournal.com]

No. That would have required him rewriting several tens of thousand lines of code. We beat him over the head with a book about real performance analysis, and he was happy.

[scs-11.livejournal.com]

952.32. Really, it's the .32 that gets me. That means their load measuring is accurate to one part in 100,000, right?

[wildw0lf.livejournal.com]

How can one user cause *that* much latency?