That's some pretty old-school thinking there. For modern storage arrays with decent caching algorithms the level of RAID protection often has little to do with IO response time and throughput. I've dealt with DBAs who have insisted their archive and redo logs live on RAID-1 or RAID-10 storage because that's what they were taught oh-so-long-ago. They want me to carve out and dedicate four spindles for them to do RAID-1 on a storage array with nearly a thousand spinning disks in it. I've got a storage array with half a terabyte of cache or more, 100% write cache hits and 95+% read cache hits. 4 spindles vs 1000 servicing IO with some extra RAID overhead? Not a difficult call.
I believe it to also be common practice to sanitize production data that goes anywhere except where it's absolutely needed. The sensitive stuff in databases gets replaced with bogus data or whacked all together. If you had, say, credit card data on various prod servers there are regulatory reasons that would prohibit a straight mirror of that data to put on a test server to play with. Not to say they follow such regulations, but it may be reasonable that a test server was compromised and nothing of value was exposed.
A little off-topic, but if you're moving from one vendor's storage virtualization approach to another's, I agree that you may need to do some sort of painful host-based migration. But if you're buying storage from multiple vendors you're basically stuck doing a lot of host-based migrations anyhow. Throwing storage virtualization into the mix means you'll only need to do this when you want to change virtualization vendors rather than every time you buy another storage array.
There is somewhat of a lock-in factor with storage virtualization, but it sure can be useful once you've got it. For completely non-disruptive data migration the work what would otherwise take months for a team of people can be scripted by one person and trickled over in a few weeks.
How many of these 2.0TB SATA drives are you going to purchase to do the same number of random cache miss IOPS that a single SSD can do? The math does not lie, applications are out there that can gain massive performance improvements and save money at the same time using SSDs. It's so easy to say hey, re-architect your application. Guess what? Mission critical apps grow organically and are not always optimized. How heavily used will your application get before even your optimized IO creeps into the realm of "I/O patterns so messed up that today's horrendous SSDs actually lower your cost per I/O"? How much money do you think the bank/nation-wide retailer/wall street firm would need to spend to "rethink their information architecture"? Not to mention power and cooling of a room full of short-stroked 2TB SATA disks vs one cabinet of SSDs.
SSD is not gaining traction simply because it's a buzzword and commands huge profit margins (both are true). It works. It solves real problems. In the right cases it saves money. If you spent some time in a larger organization I suspect you'd change your tune. You're comparing 2TB SATA apples to 256GB SSD oranges. Both may be fruit, but they're not interchangeable.
How many multi-petabyte enterprise data centers have you seen running SSDs as their primary storage? None. Yeah, that's what I thought.
Agreed that SSDs have a long way to go on price to compete, but it's simply not true that they're not yet ready for the enterprise datacenter. All the larger enterprise storage array vendors (EMC, HDS, IBM, NetApp) say they're ready, and most are shipping them with decent sales. Despite their price and the "fact" you've so eloquently stated, you'll find them in many Fortune 500 datacenters simply because they outperform spinning disks by such a factor that they're cheaper per IO. I believe today the vast majority of vendors providing enterprise-class SSD drives are sourcing them from STEC. They play some tricks to work around write limits, but they've got ~5 year MTBF ratings.
In reality, it's probably not as much of an IQ test of a potential employer as it is a test of potential employers' first line of HR flunkies. I've dealt with some painfully inept recruiters/HR people for otherwise respectable companies. You may be missing out on a good employment opportunity instead of dodging the bullet you refer to.
I'm in an inner ring suburb of Minneapolis, and I have 4 white oaks in my yard. A couple of these hang over my story and a half home, and every fall my wife and I are used to hearing them drop on our roof at night, tumble down the slope and then *tink* hit the gutter. It keeps you up some nights. Ever since we bought the house 5 years ago I've had to collect all these off the patio and out of the gutters.
1.4TB/s, wowza that's fast. Back of the napkin math says your 138 disk HP 5000 series SAN must have well over 3000 outward facing 4Gbps FC ports. Don't think so. Not sure where your number came from there.
Things may have changed some since you looked into Exchange for large environments. I work in an environment where Microsoft is pushing direct attached storage over SAN for a 200k+ user Exchange implementation. They even say they will handle remote replication this way. Microsoft has completely changed their tune regarding SAN for Exchange 2007.
Maybe they are rebranded Qlogic switches...I recall the old 9020's were Qlogic underneath (Cisco's first 4Gbps FC switches). Those things didn't understand VSANs when they came out, which seemed like a strange thing to put into the MDS line. I guess they just wanted to get out there with a 4Gbps switch.
There's something to be said for port density in large switches vs. small edges everywhere. Where I work we use Cisco 9513s in the cores and at the edges these days. We're migrating from 20+ physical McData switch fabrics down to 5 physical switch Cisco fabrics, and vastly increasing port count while we're at it. I never did the power calculations, I wonder what the difference per port would look like.
It looks to me like everything on your list about the 4900s could be achieved using the stackable MDS 9134 switches. You get a 64 port switch in 2RUs, 4Gb line rate ports (no oversub), hitless firmware upgrades and less power than your old 9216s. There aren't two supervisors like in the director class MDS switches, but I suspect the same is true for your Brocade 4900s (I've never looked into them).
Interesting you point out a Sun Infiniband switch as the a possible option to "merge it all together". Cisco's idea of a unified datacenter fabric is based on Ethernet, see Nexus. I dunno...Infiniband is certainly cool stuff, but could it ever overtake Ethernet in the datacenter?
By the way, it may sound like I work for Cisco. I don't. I do, however, manage large Cisco MDS Fibre Channel SANs.
Didn't you buy Cisco Fibre Channel gear then? I don't recall their FC gear being inexpensive either!
The 10Gb ports aren't really about the hosts (today anyhow). They're generally more useful for the connections to large storage arrays which can push that kind of bandwidth, you'd be able to fan in more hosts to each storage array port.
Kind of. This method of aggregating bandwidth by using multiple links does poor man's load balancing. The traffic between one source and one target will only traverse a single path until that path fails. If you have a lot of different sources on one side of an etherchannel going to a lot of different targets on the other side of the etherchannel, you get a relatively balanced workload. If you've got a smaller number of sources and targets it's easy to get uneven bandwidth utilization across those links. This is more common in networked storage than in everyday IP land.
Well, I specifically asked this question of an Equallogic salesman at the bar last year and he told me this was how it worked. He may have been blowing smoke or not understood the question. Either way, he was buying:).
Somebody already replied and said that what you're suggesting doesn't scale (they're right), but to add to that, here's why. When people invest in SANs and storage arrays, they typically want to take advantage of economies of scale and maximize utilization. That means they aren't going to put one application on the box and tune for it as you suggested, but they'll be putting 5, 20, or 100 applications on the same array. It's nearly impossible to tune each individual application's storage without dedicating spindles to each (or groups of each). Once you've done that, you may as well be buying separate storage for each application, and you're back to direct-attached storage without the economies and utilization you were after in the first place.
Striping all data across as many spindles as possible ensures reasonably predictable performance, and almost always better performance than you would have had by hand-tuning.
Well, saying laptop drives isn't really fair. They've been making 10,000 RPM serial attached SCSI disks in a 2.5" form factor for quite some time, I know Sun uses them in servers. I'm not sure if there are 15,000 RPM disks out yet in this size or not. These are not your 5400RPM laptop drives.
Off the top of my head, all of the following companies have storage arrays which basically do exactly what you're asking for. When you create a LUN it's across all available spindles and data will re-balance across all available disks as you add more, all with RAID redundancy. I'm not sure about N+2 at this point, but RAID-6 is becoming ubiquitous in the storage industry.
HP (EVA)
3Par
Dell/Equallogic
Compellent
Pillar
HDS (USP)
I'm not going to educate you except to tell you to Google for it.
The disruptive part is that it seems to be much more reliable which would mean that you can wave the tech goodbye for a while, instead of having to lose access to a sting of drives RAIDed together while they have to rebuild a drive which failed and needed replacement.
Umm...I think you forgot what the "R" in RAID stands for. You may have somewhat degraded performance during a rebuild when you spare in for a drive which has failed, but you don't lose access to any data because of a single disk failure (Save for RAID 0, which isn't really RAID to begin with).
I wouldn't call this disruptive. It sounds like they've done some smart things to bring disks back to life when other hardware would call them failed, but you can bet that they're packaging more spares in these non-user serviceable enclosures than you would in a user serviceable configuration.
Slashdot Mirror
GopherVR?
Here's one, Cumulus Networks. A lot of Cisco switching gear is Linux underneath with a more familiar Cisco CLI.
That's some pretty old-school thinking there. For modern storage arrays with decent caching algorithms the level of RAID protection often has little to do with IO response time and throughput. I've dealt with DBAs who have insisted their archive and redo logs live on RAID-1 or RAID-10 storage because that's what they were taught oh-so-long-ago. They want me to carve out and dedicate four spindles for them to do RAID-1 on a storage array with nearly a thousand spinning disks in it. I've got a storage array with half a terabyte of cache or more, 100% write cache hits and 95+% read cache hits. 4 spindles vs 1000 servicing IO with some extra RAID overhead? Not a difficult call.
Maybe you should quit Initech and apply at Intertrode?
I believe it to also be common practice to sanitize production data that goes anywhere except where it's absolutely needed. The sensitive stuff in databases gets replaced with bogus data or whacked all together. If you had, say, credit card data on various prod servers there are regulatory reasons that would prohibit a straight mirror of that data to put on a test server to play with. Not to say they follow such regulations, but it may be reasonable that a test server was compromised and nothing of value was exposed.
Whamcloud? Really?
There is somewhat of a lock-in factor with storage virtualization, but it sure can be useful once you've got it. For completely non-disruptive data migration the work what would otherwise take months for a team of people can be scripted by one person and trickled over in a few weeks.
It's the same as an 'old school' monkey. You can recognize them by their gold chains and Adidas.
We're all irreplaceable until that time when someone devises a machine/process/alternative
Or shell script...
SSD is not gaining traction simply because it's a buzzword and commands huge profit margins (both are true). It works. It solves real problems. In the right cases it saves money. If you spent some time in a larger organization I suspect you'd change your tune. You're comparing 2TB SATA apples to 256GB SSD oranges. Both may be fruit, but they're not interchangeable.
Agreed that SSDs have a long way to go on price to compete, but it's simply not true that they're not yet ready for the enterprise datacenter. All the larger enterprise storage array vendors (EMC, HDS, IBM, NetApp) say they're ready, and most are shipping them with decent sales. Despite their price and the "fact" you've so eloquently stated, you'll find them in many Fortune 500 datacenters simply because they outperform spinning disks by such a factor that they're cheaper per IO. I believe today the vast majority of vendors providing enterprise-class SSD drives are sourcing them from STEC. They play some tricks to work around write limits, but they've got ~5 year MTBF ratings.
Yes, sorry to say it was in fact all for nut.
In reality, it's probably not as much of an IQ test of a potential employer as it is a test of potential employers' first line of HR flunkies. I've dealt with some painfully inept recruiters/HR people for otherwise respectable companies. You may be missing out on a good employment opportunity instead of dodging the bullet you refer to.
This year -- Zero. Not a one.
1.4TB/s, wowza that's fast. Back of the napkin math says your 138 disk HP 5000 series SAN must have well over 3000 outward facing 4Gbps FC ports. Don't think so. Not sure where your number came from there.
Things may have changed some since you looked into Exchange for large environments. I work in an environment where Microsoft is pushing direct attached storage over SAN for a 200k+ user Exchange implementation. They even say they will handle remote replication this way. Microsoft has completely changed their tune regarding SAN for Exchange 2007.
There's something to be said for port density in large switches vs. small edges everywhere. Where I work we use Cisco 9513s in the cores and at the edges these days. We're migrating from 20+ physical McData switch fabrics down to 5 physical switch Cisco fabrics, and vastly increasing port count while we're at it. I never did the power calculations, I wonder what the difference per port would look like.
Interesting you point out a Sun Infiniband switch as the a possible option to "merge it all together". Cisco's idea of a unified datacenter fabric is based on Ethernet, see Nexus. I dunno...Infiniband is certainly cool stuff, but could it ever overtake Ethernet in the datacenter?
By the way, it may sound like I work for Cisco. I don't. I do, however, manage large Cisco MDS Fibre Channel SANs.
The 10Gb ports aren't really about the hosts (today anyhow). They're generally more useful for the connections to large storage arrays which can push that kind of bandwidth, you'd be able to fan in more hosts to each storage array port.
Kind of. This method of aggregating bandwidth by using multiple links does poor man's load balancing. The traffic between one source and one target will only traverse a single path until that path fails. If you have a lot of different sources on one side of an etherchannel going to a lot of different targets on the other side of the etherchannel, you get a relatively balanced workload. If you've got a smaller number of sources and targets it's easy to get uneven bandwidth utilization across those links. This is more common in networked storage than in everyday IP land.
Well, I specifically asked this question of an Equallogic salesman at the bar last year and he told me this was how it worked. He may have been blowing smoke or not understood the question. Either way, he was buying :).
Striping all data across as many spindles as possible ensures reasonably predictable performance, and almost always better performance than you would have had by hand-tuning.
Well, saying laptop drives isn't really fair. They've been making 10,000 RPM serial attached SCSI disks in a 2.5" form factor for quite some time, I know Sun uses them in servers. I'm not sure if there are 15,000 RPM disks out yet in this size or not. These are not your 5400RPM laptop drives.
HP (EVA)
3Par
Dell/Equallogic
Compellent
Pillar
HDS (USP)
I'd be shocked if Xiotech doesn't do this today.
Umm...I think you forgot what the "R" in RAID stands for. You may have somewhat degraded performance during a rebuild when you spare in for a drive which has failed, but you don't lose access to any data because of a single disk failure (Save for RAID 0, which isn't really RAID to begin with).
I wouldn't call this disruptive. It sounds like they've done some smart things to bring disks back to life when other hardware would call them failed, but you can bet that they're packaging more spares in these non-user serviceable enclosures than you would in a user serviceable configuration.