Slashdot Mirror

I am going to respond because you are just a moron! Not total per rack TOTAL! Lets say you have 100 servers and they are in 10 RACKS. You consolidate these 100 Servers in 10 racks through virtualization. You now have 10 servers in 1 rack, carrying the load of what was 100 servers in 10 racks. The total power for that 1 rack will go up! The total power overall will go down, EVERY SINGLE TIME. You win the award for the densest person on the Internet. I have to assume you are just fucking with me at this point because no one who can type can be as idiotic as you. (I gave you an out , you might want to take it).

AC Costs absolutely become different! Some systems can not handle the higher density BTU load properly! But who cares because your assertion all along is that total power consumption goes up, not "Your current cooling system might need to be redesigned to accommodate the denser heat output". If you would like to discuss the impact of virtualization on Data Center cooling , I would be happy to. First you need to let us all know that you have finally grasped the fact that total power consumption does not go up. Until you can grasp that fact there is no point in trying to help you understand anything else.

The fact you think you need an additional cooling system proves you do not understand data center cooling either, 9/10 times the problem can be addressed with proper planning of the rack layout and its relationship to the existing CRAC units. Want to know why? Because Total BTUs went down right along with the Total Power. You of course do not believe using less power overall actually is a reduction in total power consumption. I am sure you will now link me some article you misread that you believe states virtualization increases the total BTUs in the datacenter.

You are just grasping at straws, trying to save some face. Man up, admit you were wrong and move on.

You like to read stuff on the Internet. You seem to focus on writers speaking hypothetically. Here are some real people talking about their real power and btu reductions:

http://virtualize.wordpress.com/2006/12/18/power-s avings-through-virtualization/

http://blogs.hds.com/david/2006/11/a_power_company s_reward.html

http://www.foedus.com/solutions/datacenterConsolid ation.php

(conslidation and virtualization are 2 different ways of skinning the same cat, you still end up with fewer servers at a higher load)

http://www.solutioncentre.co.uk/blog/index.php?ent ry=entry070517-222751

It all depends on what you are trying to do

For some workloads, many servers with four drives each may work. This is the Petabox/Google model. This works if you have a parallelisable problem and can push most of your computation out to the storage servers.
Remember, you don't have a 300Tb drive, you have 300 servers, each with 1Tb of local storage.

For other workloads, you need a big disk aray and SAN, probably from Hitachi, Sun, or HP. This is the traditional model. Use this if you need a really big central storage pool or really high throughput.
Many SAN arrays can scale into the PB range without too much trouble.

Nowadays, a PB is enough to arch your eyebrows, but otherwise not that amazing. It seemes that commercial storage leads home storage by 1000. When home users had 1GB drives, 1TB was amazing. Now that some home users have 1TB, many companies have 1PB.

If it's cheap, then Netapp might not qualify... :)

What about technologies - NAS? Host-attached? Gateway/NAS? Grids?

Other companies/products to consider:

EMC (The Celerra is a nice product)

Onstor Bobcat

HP

IBM

Hitachi

Panasas

Exanet

Yotta Yotta

StoreAge

If you want basic raid devices look at Infortrend/Transtec. Their S-ATA offerings now support RAID-6 and are dirt cheap.

You rang Milady?

Some of us don't get out of bed for less than a Petabyte ...

http://www.hds.com/products_services/universal_sto rage_platform/

I have six. You have to love 2Gb fibre channel. I like using RAIDSilly (Redundant Array of Independent Datacentres) ...

Microsoft offers this as a solution to long-distance fault tolerance when used in conjunction with the iSCSI protocol (SCSI over IP).

"WAN" clustering/majority node set clustering is possible with SAN replication as well. Not cheap, but nice.

http://www.hds.com/pdf/wp179_storage_cluster_windo ws_2003.pdf#view=FitH&pagemode=bookmarks

Gimmick? Hmm...

This is just electronic writes. Those who have worked with a high-performance SAN like the Hitachi 9900, Sun 6920, know that electronic writes is where all the performance comes from. When our SAN's 4gig of cache goes offline, my DBAs come running and everyone complains about terrible write/read speeds.

Electronic writes (in a good amount) means that the data flies into memory and later on the disk system pumps the data out to the disk platters. Netapp is really great at doing this kind of operation. Talk to someone with a Netapp and they'll tell you that you can watch all the activity lights on an array light up in a fury of activity with the Netapp head pushes from cache->disk. It's really fantastic. Of course, it's really expensive too.

I'm not saying that Samsung's disks are going to be enterprise-class. But I'm hoping that this cache thing catches on to make SATA2 (or whatever) based arrays feasible for desktops. Imagine a chain of large-cache drives on desktops, large-cache single drives on laptops ... For example, if you had 128MB of cache per drive, then an array of 8 disks would have 1gig of cache. That would be so great for the power-user. The Sun T4 bricks cost $50k and they have 2gb of cache per controller. 2gb of electronic writes/read = amazing performance.

I can't wait to learn more about the quality and speed of the disk cache. It's what I've always wanted to see in consumer grade disks.

I'm not entirely sure, but I believe that the credit for the Deskstar goes entirely to IBM, and shares the Hitachi name only because of the JV.

I *do* know that HDS makes some of the nicest hardware I've ever seen (right down to the drives), and I would not hesitate to buy a Hitachi-manufactured (as opposed to labeled) drive.

For an example of how Hitachi gets it, see this page:

http://www.hds.com/products_services/storage_syste ms/enterprise_storage/

I've spent a chunk of time lately playing with a Sun/Hitachi 9980. Imagine a fiber channel array of hard drives the size of a nice, hefty subzero 2-door refrigerator (2m x 2m x 1m, roughly, for 1 control module and 1 array module).

It hooks up to a dozen computers, has room for over 100TB of drivespace (raid-5), has an configuration console beyond the OS that allows some slick on-the-fly tricks, is compatible with virtually ANY OS, lets you slice the array a zillion ways, gives you a data pipe of Gigs per second, and costs a million dollars. Now that's some serious power: you could capture the entire speex-quality audio of 400 people's entire 80-year lives on it (400 x 80 x 365.25 x 24 x 3600 x 1k/s = 1.0098 x 10^15 bytes, or 100TB).

But... one day I was trying to find words for how cool this thing is, and I realized: I can remember paying a buck a byte for memory, and wincing at HD prices. I also still have a ST225, for nostalgia or whatever reason. And a 250gig drive is down around $100 now, so I'm just 2^9 away from 100TB. A conservative pseudo-Moore's law rate for HD's gets me there in 20 years: my ST225 (20Meg) is about 20 yrs. old, or 2^13 in 20 yrs).

Given the exponential rate of storage growth, I am less than 20 years from being able to buy one of these puppies at commodity prices. And by 2030 it'll fit on my wrist.

EX-cellent...

Hitachi HDS has some nifty SAN products/technology. Merge that with some of IBMs SAN (think ESS) technology and you got yourself an interesting outcome. Come to think of it, IBM and HDS have been in bed for awhile now. This seems more of a formality and merging of technologies to compete againts EMC.

IBM and Hitachi are *merging* their disk business so that IBM gets a 30% stake (and Hitachi, 70%). The story's comment "They plan to sell 70% of the their HD business to Hitachi." seems incorrect to me; IBM is simply estimating that its current disk business is worth 30% of the joint disk business. Also, note that Hitachi has a very strong storage systems business HDS (right behind EMC) that is very profitable (also resold by SUN as Storedge9900 series datacenter/enterprise storage products, I believe), so big blue may have merged their disk business with a view to ensuring future profitability in the overall storage space.

As best I can tell, please tell me if I'm wrong but that is almost exactly what EMC does for their NAS products. The processors are PowerPC, I think, but they use their own OS and controlling software. Hitachi has an excellent Storage primer, its very basic and written for nontechnical stock analysts (in pdf) here. Unfortunatly it hangs the conversion tool. It also crashed at work under Winn 98 (fine under Win2k) when printing at about page 185. The link is to the file which is quite large (about 6 megs). I found it to be rather useful but somewhat basic. Also, EMC usually charges at least $100k for their wheel. Veritas sells just the software. I would doubt its cheap, but probably less than writing your own comparable product.

You are correct that you should be doing analysis of your system to identify bottlenecks and then work to eliminate them. However, this should also include the application level, as well.


I can't speak for how one would go about doing application-specific tuning for things specific to Exim, but I find that many MTAs share the same problems. I'd suggest going to the Sendmail Performance Tuning for Large Sites paper that I wrote and presented at SANE'98, and see what of those problems (and solutions) might be applicable to your situation.

It is entirely possible that you could end up tuning the system performance enough that you don't even need to buy any additional hardware at all, just change the configuration of the software and OS you already have.


That said, if you've done all this and you still have problems, you probably do need to buy some new hardware. If that new hardware you need to buy turns out to be disk storage, Sun will be glad to sell you a StorEdge system that will implement RAID levels 0, 1, and 5 in hardware.

This would probably be the simplest and easiest solution to implement on a Sun machine, since the people who sold and installed the machine originally can help you with the expansion. However, it's likely to be rather expensive from a price/performance perspective. Note that Sun OEMs their hardware from SymBios/LSI Logic, and you can buy higher-end equipment direct. See http://www.metastor.com/ for more info.

If even the MetaStor hardware isn't enough for you, then you might want to consider vendors such as EMC and comparable units from Hitachi Data Systems. For my part, the HDS equipment can have a larger cache (up to 16GB in some units), can segment the cache so that different hosts get their own dedicated slice (which EMC can't do), and overall seems to simply be more intelligently implemented.

If you were in Europe, I would suggest looking at Comparex, since they are the licensed HDS VAR for this region.

1. It doesn't scale well above 2 CPU's
2. It is HORRIBLY UNPREDICTABLY UNRELIABLE!!!
3. It is very hard to remotely manage, so every time you have a problem you might have to run to the data center to even do troubleshooting, then you find that you need that diskette in your office..., then you need to make a phone call, but the data center phone is 20ft away from the console (I HATE NT IF FOR THIS REASON ONLY!).