Written by W. Curtis Preston
Thursday, 20 August 2009 10:56
My good friend, Steve Duplessie, wrote a
blog article that basically said that the issue that dedupe was designed to solve is "NO LONGER VALID" (Caps his). He didn't say dedupe is bad, but he says that he can buy JBOD for 1/9th the cost of a deduped system; "therefore, using deduplication to solve an economic scarcity issue is no longer legitimate." He also said that "if I'm off by 100%, I'm still 1/4 the cost."
With all due respect to my good friend, I don't agree with him any more on this than I do on whether or not you should root for the Patriots or the Chargers. His logic is sound, but his numbers are off -- off by a whole lot more than 100%. And there's another "scarcity" factor that he's not considering.
Steve's
post is a follow-on to
another post of his that was inspired by a book called "
Free" by Chris Anderson. Among other things, this book apparently talks about the concept of "scarcity" and how things that solve scarcity problems make money. He says, and I agree, that backing up to disk was to solve the scarcity of time when backing up to tape. Then dedupe was made to solve the scarcity of money, since backing up to disk costed too much. But then he says that this is no longer an issue because cheap disk costs so much less now, and he makes his point by comparing the alleged cost of a 30 TB Data Domain system (which he says costs $90K) to a 30 TB JBOD system (which he says costs $10K). He's saying therefore the issue that dedupe was trying to solve is no longer valid.
First, let's talk the Data Domain pricing used in the blog post, as it definitely doesn't match what I'm seeing. I verified today that the list price of a Data Domain system capable of holding 30 TB of backups is $32K -- not $90K. That's for a DD510 with one expansion tray that gives you 2.7 TB of usable capacity. With an 11:1 dedupe ratio, which is a realistic ratio, that will hold 30 TB of backups. Just for comparison, Quantum also has a 1.9 TB system that lists for $12.5K. Two of those and a 10:1 dedupe ratio and you've got yourself 40 TB of backup capacity for $25K. Taking out the additional capacity makes the effective list for 30 TB of Quantum about $19K ($25K * 30 / 40).
As to the JBOD system, I have no idea where I'm supposed to buy a 30 TB disk array for $10K. Let's look at a few disk systems. A Dell
MD 3000 configured with 15 1 TB drives sells direct for $16.5K, so that's $33K for 30 raw TB (pre RAID5). But I need 30 TB usable, which means I'll need to buy three of these for a total cost of $50K.
If you're OK with not buying a big brand name, the least expensive arrays I've seen in the middle enterprise are NexSAN arrays, and they list for about $30,000 for 30 TB -- raw. A reseller that I know says that their rule of thumb (based on the way customers usually configure them) puts them about $40K-45K for 30 usable TB.
OK, forget any kind of brand name and let's just go for cheap. The least expensive arrays that I've heard of (but never seen in a customer) are Promise arrays. A
Promise VTE310F array sells (direct) for $6789 with no hard drives. Filling it with 12 1 TB disk drives from
buy.com for $90 apiece adds $1080. So that's $7869 for 12 raw TB using the cheapest arrays and disks I can find. I would need three of them to get to 30 usable TB, costing $23K.
It's also important to point out that there are many classes of storage in the enterprise, and comparing a Promise array to a Data Domain or Quantum system is ignoring those classes completely.
So, that's $19K-32K
list price for the dedupe option and $23K-$50K
street price for the JBOD option. Without looking at anything else, deduped storage is cheaper than JBOD disk -- even if you build it yourself.
And that's just the acquisition price. Remember that the deduped system will be using at least 10 times less power and cooling than the JBOD system, and that's a huge deal. Steve said that power isn't scarce and hasn't been for a long time. Are you kidding me? I've been in datacenters that have been told by the power company that they can't have any more power. And, even when it's not scarce, it's expensive -- and anything we can do to reduce it is a good (and a green) thing.
Secondly, deduping backups enables replicating those backups, which simply isn't possible in most datacenters without dedupe. Trust me -- bandwidth is scarce. Now you can have onsite and offsite disk-based backups without moving tapes around. If you want to make tape copies, make them offsite -- and they never get moved.
Finally, Steve's post missed another important point. Disk is not in competition with deduped disk; it's a component of deduped disk. So as disk gets cheaper, so does deduped disk. As disk prices have fallen over the last several years, so has the per-GB pricing of dedupe systems.
So even if it costed more (which it doesn't), then it would still be a Good Idea for those reasons. And that's all I have to say about that.
I like Steve a lot, but I think he needs to check his pricing numbers again.
Add comment
Comments
Having said that, you never heard me say that dedupe was a way to save money when compared to backing up to tape. It's simply BETTER than backing up to tape. Backups are easier, restores are easier (and yes, typically faster), and management is easier with disk than tape. I never said it was cheaper than tape.
It IS cheaper than backing up to disk without it, which was the original point of this post.
To get 900MB/s of restore speed I have to invest at least 400K to purchase a single 880.
To get the same speed from LTO-4 I would need 6 drives, or maybe less. keep in mind that my environment which is probably not typical to what you see based on your description is designed from the ground up for fast recoveries. For these tapes with a nice high speed robot and 100 slots will cost me probably around the 200K. you know what, I'll be generous with you and add a disk storage system for those clients who can't push data fast enough for another 100K. Where is my savings with de-duplication ? you are right, on the paper.
Disk is easier ? OK. For De-duplication appliances in VTL mode you can not use multiplexing, so you will have to create a lot of virtual tapes to satisfy your performance requirements, a management nightmare for any recovery application.
For the NAS mode, multiplexing is not a problem, for Legato at least, but then you are hitting a size limit that will eventually affect your cloning performance, you put it well, single stream, it takes 12 hours to move 6TB of data to a single LTO-4, you want to cut the time ? create and manage multiple NAS shares, yes, disk is sooooo easy.
I will finish this discussion with one of my favorite quotes: "I would rather die of thirst than drink form a cup of mediocrity - Stella Artois"
The Doctor
You are generalizing again. Not all of them even HAVE a database or anything like it that is used during restores. In addition, having one database/hash table/or anything else is not bad as long as it can respond to all the requests.
And I think you're misunderstanding the single stream concept. It doesn't mean that other restores have to wait. The question is what is the fastest a single restore/copy can go if it only has one stream. Copies to tape by definition only have one stream, and often large restores also only have one stream (e.g. one large backup of one large filesystem). This number is important because an individual restore/copy can never go any faster. Once you know the single stream restore speed, you then want to know how many copies/restores of that speed the appliance can support. For example, a DD 880 supports a single stream restore speed of just over 300 MB/s, and it supports an aggregate speed of 900 MB/s for non-NetBackup backups, so it should support three such restores before individual restore speed starts degrading.
In real life, restores rarely approach the maximum single stream restore speed of most appliances, as restores are usually throttled elsewhere (as you mentioned). But if you do need a really fast restore and you can design your infrastructure to support it, it's important to know the single stream performance of the appliance(s) you're considering.
And of course I can help them get better with their tape drives. Currently my best advice is to go to disk first. It's sooooo much easier.
As for single stream, couldn't agree more, however, for best possible recovery performance you must have a dedicated, noise free, channel from your recovery source to your target. Problem with a de-dupe appliance is it has only one database that keeps track of all the mess on the disk, and I don't think you would dedicate an entire appliance just to recover a single application. in disaster recovery, time is crucial and the pressure is high, I mentioned real life twists, think about "hey Curtis your wife just called, the storm that destroyed your data center is destroying your house and your cat is trapped in the basement", in your example a possible response will be " wait honey, I have to wait for my SAP to be recovered, since single stream is the king, so I could start my Exchange recovery later,that will take a few hours so in the meantime I could come home and help evacuate you and fluffy".
Tape or traditional disk source will be much more efficient in that case. A good DR will take 25% longer than your last full backup window, and this is a hard proven repeatable fact in my organization. and for a DR to be successful the secret is do more, wait less.
you say 90% can't meet their tape's MTS ? I want to believe that is the real problem, I thought "Mr. Backup" is seasoned enough to help those people solve this one and I really hope that your solution will not be, change your backup paradigm, suffer in recovery and throw more money on licenses and maintenance.
One more thing, when a de-dup appliance that answers all the items you mentioned will be available and proven mature, I will be the first in line to buy it. Meanwhile, don't waste our time. The last sales guy who tried to pitch me a de-dup product, now has to sing is ave maria song with a lot of faith and dedication if he ever wants to see me again.
If a dedupe product does this:
1. has fast backups (all types)
2. gets my stuff offsite quicker and safer than tape (no encryption required)
3. can restore my stuff faster than tape (I know this is your sticking point)
4. can restore my stuff at close to the speed of non-deduped disk
5. has a similar cost to tape and disk (see the post this comment is in)
Then why wouldn't you want to buy it? You're saying that it can't meet those requirements, and particularly can't meet #3 or #4. And I'm saying that I have seen systems that can. I've also seen systems that fail miserably at all of them. But the majority that I've seen are better are most all of those than tape is.
You recommended just buying a tape drive or two. If people could actually keep an LTO-n tape drive happy, I'd be fine with that. But 90% of customers (that's a hard number) that I have visited are getting less than half the rated throughput of their tape drives, and 50% (again, hard number) are getting less than 20% of the rated throughput. And that problem only gets worse as drives get faster. And the worse it gets the worse backups get. Disk is the perfect target for backups, and dedupe lets us keep it on disk longer -- and a good dedupe system DOES NOT slow down restores.
As to testing, I believe that if you're going to store 90 days worth of backups in production, you should do that in test. But it doesn't take 90 days to put 90 days of backups into a target dedupe system if you automate it using tape copies. You can usually do such a test in a fraction of real time.
Glad to hear MS is getting on the bandwagon that NetApp & VMware started over 10 years ago.
BTW, the appliance from the storage giant who I was referring to in my last post, did perform at 100GB/Hr (~30MB/s).
But I believe, Curtis, the basic question still remains, is de-duplication is really an adequate technology from a price/recovery performance perspective and the answer behind the picture you are painting is still no, in my opinion.
Any financial model you put on the table just don't fit in and any "imaginary" savings/ROI that occur when deploying this technology can not be compared to the extended downtime your business will suffer during a real disaster.
I second you that everyone should test, retest and then test again the technology before deploying it, that's why a POC in my book is no less than 30 days of extensive performance/scalability testings, in most cases, I believe, you will find that the better investment is in the little things that makes a big difference when time comes.
Looking ahead, I really like what Microsoft is doing with VSS, especially block level incremental transfers, I heard that VMWare are working on doing the same, if all recovery application vendors will adapt this it is going to be a wonderful world to live in.
Have a great weekend !
The Doctor
BTW, takes a big man to back down. Good on ya'.
It was 100MB/Sec (360GB/Hr) on average.
The vendor who told me that this is the cost of doing de-duplication is the same vendor that owns my recovery application, and that was before the latest acquisition.
For customers who run backups at 150MB/s (~1 X LTO-4), still cheaper buying 2 drives and a lot of tapes than buying a single de-dup appliance so what are you saying, Curtis ? that de-duplication is the most expensive poor man solution for backup ?
The Doctor
I again ask that you contact me offline so I can understand more of what you're trying to say. My email is curtis - at - backupcentral.com.
You say that what he wrote proves your point. I don't follow. You said:
Quote:
100 GB/Hr is 27 MB/s. He posted that he's getting five times that, and you say it proves your point.
You seem to think it proves your point because you're comparing his 150 MB/s to your 2 TB/hr number. You can't do that because you don't know what his backup speed is. His boxes are a few years old. A few years ago there was no such thing as a target dedupe box that could go 2 TB/hr. In fact, Data Domain didn't ship such a box until last month. And until they upgraded their code several months ago, their best number was about 1 TB/hr. You also don't know (because he didn't say in his post) whether his 150 MB/s is single stream or aggregate number. So I'm asking him for clarification on that.
But my summary is that you said that dedupe was "a disaster for recovery" because restore speeds were never faster than 27 MB/s, and that definitely is not the case. It's not the case in my experience, or in the experience of this user. Again, I'm not saying that this didn't happen in your testing. I'm saying it's not happening in other people's testing.
RSS feed for comments to this post