So you need about 30 AMD machines to get the same speed. That's about $250k including rack and networking. Right off the bat you're talking about a 1/8 the cost of the IBM power system.
As for performance/watt, the AMD machines need about 600W each. A rack plus switch is probably going to need 20kW to run.
Of course I'm having a very hard time trying to find the power requirements for a fully loaded 795.
10ms comes from the paper I read on the average page load time (latency) increase for https vs http.
Amazon's page load time is not limited by https. They have far worse problems for first load. I used http://analyze.websiteoptimization.com/ to pull down www.amazon.com.
Total HTTP Requests: 69 Total Size: 445054 bytes
445kbytes is going to take several seconds just to download if you get 1mbps to their CDN servers.
To answer your original question, yes, I know how performance oriented big sites are. I work for one.
Wrong, There are many papers out there showing that https is not a latency hit compared to page download time and server side rendering. From memory it was maybe 10ms of time for new connections and 3-5ms for SSL resumes. You only take a real CPU hit on the first connection from a user.
Google can handle all of gmail over https. Think about how many ajax hits that is.
Most sites can't even classify their latency requirements properly let alone figure out if https is "too expensive"
The problem with DC in the server room is distribution. You would have to go high voltage DC to be able to distribute it.. and then you would need to step it down a bunch at the rack. If you do 48v rail, you're going to need huge distribution bars. Think of a typical server rack these days. A 500W 1U server x 40U. That's 20kw per rack. At 48v you're talking 416 Amps. A minor install of say 10 racks is going to bring you up to 4160 Amps. Just guessing you'd need a bar of copper 2x2" in size to move that kind of current.
Considering modern server AC/DC power supplies are 95+% efficient, and common server distribution is done at 208v, there's no need to go DC at the datacenter level.
Absolutely true. Every time I watch traffic from my commuter bus on the highway I feel very very glad I'm not out there driving.
It gets worse when I'm out cycling. I was cycling on a nice quiet park road in SF (Washington+Arguello Blvd) the other weekend and some tourists were driving 15mph swerving in and out of the shoulder/bike lane. I ended up picking a safe time to pull around and pass them on the left since I wasn't interested in waiting for them to do something stupid like slam on the brakes.
Hell yes. Robot machine replacement would rock. Just like tape swapping. Just bring the busted machines to techs for repair. Or if it's just drive swaps the robot could do that too.
In theory you could use waste heat as a warming plant for other housing/office use. Since most hardware is happy with 80 degree input air you would have near 80 degree output water. That's more than sufficient to pipe to warm up homes in cold areas of the world.
100 degree hot isles are too cold. Hot isles should be the temperature near the maximum component tolerance of the parts in the server. If a part has a maximum temperature of 150 degrees, and runs happily at 120 degrees, the hot isle should be 120 degrees. This way the cooling efficiency is the highest.
I work on large cluster computing systems. I deal with this every day at work. One machine that's doing strange things like causing every third job to SIGSEGV is annoying and I take it out of production, wipe it, run it through memory and CPU tests and then put it back. Of course this work is not really something I have to think about, I just flag it and automation takes over.
My real job is when this machine comes back from testing still broken. I dig in, find out what is wrong.. could be CPU, memory, some other random hardware defect on the mainboard. Once it's root caused and testable the test can be added to the automation and I don't have to do anything but whack-a-mole for similar problems in the future.
And that's just one aspect of what's going on day to day.
I read the comic around the same time as the movie coming out. I personally hated the squid ending. It was much more in line with "the smartest man in the world" to simply manipulate Dr Manhattan into being the scapegoat.
Unfortunately you can't use a Yubikey for gmail right now because it is using TOTP not HOTP. Yubico would have to make a battery version of the token to support the clock needed for TOTP.
The Google Authenticator project includes implementations of one-time passcode generators for several mobile platforms, as well as a pluggable authentication module (PAM). One-time passcodes are generated using open standards developed by the Initiative for Open Authentication (OATH) (which is unrelated to OAuth).
These implementations support the HMAC-Based One-time Password (HOTP) algorithm specified in RFC 4226 and the Time-based One-time Password (TOTP) algorithm currently in draft.
Unforunately TOTP is still in draft. But you're welcome to implement your own TOTP token app for $DEVICE, or buy a TOTP device and import the key when signing up for 2-Step.
Again, what scale? Enabling https is only a few % different in CPU time for handling the crypto overhead. I've done the math. Based on any reasonably modern server machine (say a 1U dual socket quad-core) and facebook's quoted query rate it would only require an extra half rack of CPUs to turn on https for all facebook pages, including images.
It's also especially easy when you require that most of your "apps" are server based. And the only thing the client really needs is a browser. Most companies larger than a few hundred employees need some kind of server infrastructure anyway to allow multiple people to access the data. Why keep using desktop apps when browser-based apps work just as well for minimally interactive things like HR and accounting tools.
Slashdot Mirror
I've done these numbers before. How about this:
IBM Power 795 (4.0 GHz, 256-core) 1TB ram - specint rate 2006 = 11,200 - $2m
AMD Opteron 6176 dual socket (2.3Ghz, 24 core) 128GB ram - specint rate 2006 = 400 - $8100
So you need about 30 AMD machines to get the same speed. That's about $250k including rack and networking. Right off the bat you're talking about a 1/8 the cost of the IBM power system.
As for performance/watt, the AMD machines need about 600W each. A rack plus switch is probably going to need 20kW to run.
Of course I'm having a very hard time trying to find the power requirements for a fully loaded 795.
10ms comes from the paper I read on the average page load time (latency) increase for https vs http.
Amazon's page load time is not limited by https. They have far worse problems for first load. I used http://analyze.websiteoptimization.com/ to pull down www.amazon.com.
Total HTTP Requests: 69
Total Size: 445054 bytes
445kbytes is going to take several seconds just to download if you get 1mbps to their CDN servers.
To answer your original question, yes, I know how performance oriented big sites are. I work for one.
Wrong, There are many papers out there showing that https is not a latency hit compared to page download time and server side rendering. From memory it was maybe 10ms of time for new connections and 3-5ms for SSL resumes. You only take a real CPU hit on the first connection from a user.
Google can handle all of gmail over https. Think about how many ajax hits that is.
Most sites can't even classify their latency requirements properly let alone figure out if https is "too expensive"
Wow, you seriously sound exactly like the people you're bashing. Egotistical raving asshole.
Sorry, I was just responding in the parent post's units. :(
The problem with DC in the server room is distribution. You would have to go high voltage DC to be able to distribute it.. and then you would need to step it down a bunch at the rack. If you do 48v rail, you're going to need huge distribution bars. Think of a typical server rack these days. A 500W 1U server x 40U. That's 20kw per rack. At 48v you're talking 416 Amps. A minor install of say 10 racks is going to bring you up to 4160 Amps. Just guessing you'd need a bar of copper 2x2" in size to move that kind of current.
Considering modern server AC/DC power supplies are 95+% efficient, and common server distribution is done at 208v, there's no need to go DC at the datacenter level.
Absolutely true. Every time I watch traffic from my commuter bus on the highway I feel very very glad I'm not out there driving.
It gets worse when I'm out cycling. I was cycling on a nice quiet park road in SF (Washington+Arguello Blvd) the other weekend and some tourists were driving 15mph swerving in and out of the shoulder/bike lane. I ended up picking a safe time to pull around and pass them on the left since I wasn't interested in waiting for them to do something stupid like slam on the brakes.
Hell yes. Robot machine replacement would rock. Just like tape swapping. Just bring the busted machines to techs for repair. Or if it's just drive swaps the robot could do that too.
In theory you could use waste heat as a warming plant for other housing/office use. Since most hardware is happy with 80 degree input air you would have near 80 degree output water. That's more than sufficient to pipe to warm up homes in cold areas of the world.
Better than home depot, Etymotic full frequency plugs are great:
http://www.etymotic.com/ephp/er20.html
I've also had some friends use things like these in datacenters:
http://www.amazon.com/dp/B00009363P
They let you hear people talk (bandpass filter) without letting the low/high noise in.
100 degree hot isles are too cold. Hot isles should be the temperature near the maximum component tolerance of the parts in the server. If a part has a maximum temperature of 150 degrees, and runs happily at 120 degrees, the hot isle should be 120 degrees. This way the cooling efficiency is the highest.
See Google and SGI (Rackable) container designs.
http://arstechnica.com/hardware/news/2009/04/the-beast-unveiled-inside-a-google-server.ars
As you can see from the photo there, all the cables in the front. No need to get behind it where the hot isle is.
True. But that kind of sysadmin work is boring. It's no better than being a janitor.
The real fun stuff these days is not just doing sysadmin work, but working on automation and monitoring that could replace 1000s of Cert mill morons.
I work on large cluster computing systems. I deal with this every day at work. One machine that's doing strange things like causing every third job to SIGSEGV is annoying and I take it out of production, wipe it, run it through memory and CPU tests and then put it back. Of course this work is not really something I have to think about, I just flag it and automation takes over.
My real job is when this machine comes back from testing still broken. I dig in, find out what is wrong.. could be CPU, memory, some other random hardware defect on the mainboard. Once it's root caused and testable the test can be added to the automation and I don't have to do anything but whack-a-mole for similar problems in the future.
And that's just one aspect of what's going on day to day.
Yes, you can thank the Puritans for that. Instead of the nude human body just being what it is. Nudity = sex = profane = sin.
I read the comic around the same time as the movie coming out. I personally hated the squid ending. It was much more in line with "the smartest man in the world" to simply manipulate Dr Manhattan into being the scapegoat.
Unfortunately you can't use a Yubikey for gmail right now because it is using TOTP not HOTP. Yubico would have to make a battery version of the token to support the clock needed for TOTP.
4: Google files extortion lawsuit against MPEG-LA
Read the source. Google is using open standards here.
http://code.google.com/p/google-authenticator/
The Google Authenticator project includes implementations of one-time passcode generators for several mobile platforms, as well as a pluggable authentication module (PAM). One-time passcodes are generated using open standards developed by the Initiative for Open Authentication (OATH) (which is unrelated to OAuth).
These implementations support the HMAC-Based One-time Password (HOTP) algorithm specified in RFC 4226 and the Time-based One-time Password (TOTP) algorithm currently in draft.
Unforunately TOTP is still in draft. But you're welcome to implement your own TOTP token app for $DEVICE, or buy a TOTP device and import the key when signing up for 2-Step.
Again, what scale? Enabling https is only a few % different in CPU time for handling the crypto overhead. I've done the math. Based on any reasonably modern server machine (say a 1U dual socket quad-core) and facebook's quoted query rate it would only require an extra half rack of CPUs to turn on https for all facebook pages, including images.
What truth? I'd like to see your evidence. Search results are not truth.
You seem to be on an emotional posting spree of destructive negativity, trolling, and just general FUD.
Maybe you should sit back, relax, and think about how you present yourself in public.
Nice self promoting to another domain you own.
You fail at defense in depth school.
SecurID is dead, see OATH.
http://www.openauthentication.org/
HOTP:
http://www.ietf.org/rfc/rfc4226.txt
http://en.wikipedia.org/wiki/HOTP
and eventually TOTP
http://en.wikipedia.org/wiki/Time-based_One-time_Password_Algorithm
Figuring out free space inside a *NIX system isn't that hard. Just because you lack algorithmic imagination doesn't mean it's difficult.
My users do this from time to time. Usually by leaving bottles of whiskey on my desk.
It's also especially easy when you require that most of your "apps" are server based. And the only thing the client really needs is a browser. Most companies larger than a few hundred employees need some kind of server infrastructure anyway to allow multiple people to access the data. Why keep using desktop apps when browser-based apps work just as well for minimally interactive things like HR and accounting tools.