I think that all of the solutions have their place - I'm not as interested in 'compression' as I am with 'deDupe' - they are separate, and not entirely equal.
The only issue I can foresee with deDupe, besides adding overhead, is chance of failure: if the data is stored in a single block, and it is lost, so are you - but that's why you have parity, so it shouldn't be any more dangerous than normal data storage (provided the software works adequately). As for overhead, having multiple individuals 'hit' the same block would cause performance issues - but I'm not so convinced: There's really no such thing as 'overwrite' in ONTAP, it's always writing to a free block - as for reading the single block that has been deDuped, even though multiple machines hitting the same block should ordinarily cause higher latencies for everyone, I would think that the simple fact that there is one block with many accesses would indicate that it would already be cached - caching it in memory should remove the read latency penalty.
Stubs carry their own penalties, btw: when the stub is seen, the filer must use the fpolicy service to communicate with the hosting system to redirect and\or obtain the actual data: this is multiple layers of network communication, memory lookup, and disk access before the client gets the requested data - this (in theory) should be slower than a self contained system in almost every case.
VTL is definitely a good solution for backups and we're looking into them now. But HSM products, and archiving products of all types, still have one common flaw: they move the data around and it must then be managed. While this would be the most likely best way of dealing with the massive amounts of data in anyone's environment, it would still be nice to keep it all in one place. That's just a dream, anyways... I'm not sure that there would be enough savings through deDupe to actually accomplish anything like this. Now on the other hand, if the filer were able to move data to cheaper ATA storage within the same system, and simply use a redirector in the filesystem (pointer through vvbn), that could prove useful: you'd have the same lookup of vvbn-to-pvbn in order to find the data, but the actual blocks would be elsewhere within the system. No additional layers of stubs, or redirection through some other HSM system.
Sorry for the long email this early in the morning,
Glenn
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Darren Soothill Sent: Tuesday, February 27, 2007 12:36 AM To: toasters@mathworks.com Subject: RE: data management through compression for Netapp
The issue with any compression technology is the latency it adds to any workload and what happens with a large number of people doing the same thing.
Would you not be better at looking at an archiving product that can archive off the older files leaving stub files behind so any application can carry on working with no interruption. Assuming that the data is stored as lots of files.
Perhaps you need to look at some traditional HSM type product that integrates with a tape library except instead of using a physical tape library you use a VTL which will do the deduplication and compression for you. You then have the added advantage that the VTL can spit out copies of the virtual tapes to be taken off site for DR. This type of solution should then be platform independent.
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Walker Sent: 27 February 2007 03:18 To: Glenn Dekhayser; Wilbur Castro; toasters@mathworks.com Subject: RE: data management through compression for Netapp
I've also heard good things about Data Domain.
On the other hand, does anyone have any experience with the A-SIS deDupe technology from NetApp? Sounds great, except for the 'out of band' nature of it.
Glenn
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Dekhayser Sent: Monday, February 26, 2007 8:36 PM To: Wilbur Castro; toasters@mathworks.com Subject: RE: data management through compression for Netapp
Wilbur:
What are your performance requirements? Adding compression/deDup into this has a serious impact on I/O performance, it will pretty much totally screw any read-ahead optimization.
If you still want to go ahead, look at Data domain's Gateway series. It doesn't formally support Netapp, but i'm sure you could use it. This will do exactly what you want, as long as you want NFS!
Glenn (the other one)
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Wilbur Castro Sent: Monday, February 26, 2007 4:55 PM To: toasters@mathworks.com Subject: data management through compression for Netapp
Hi toasters,
We have a couple of 100 TBs of heterogenous storage (some from netapp), and see that our storage grows close to 60% over year. We were looking at alternatives for managing this data growth. Compression was one of the techniques we were considering for our nearline and (possibly) primary storage. Our applications cannot change to do their own compression, so it boils down to doing this in the storage layer or through an external device. Also, we'd like to not have any performance impact and compression to happen transparently. Deduplication technology from storage vendors would help, but it is not a hetrogenous solution.
I am not aware of any compression technology from netapp. Are you folks aware of any solutions? Would love to hear your experience with those or other alternative ways you deal with the storage growth problem while managing costs.
Thx, Wilbur
Ah, Glenn, you're so spoiled with Netapp! :-)
Problem is, with things like Data Domain, unless you're going to use the Netapp as back-end FC storage (which isn't supported yet), we're not in WAFL-land, so in most cases you are writing to the same place. And without Raid-DP, a single bit error on a raid rebuild could have a much bigger impact than normal, if that particular bit belongs to a byte that's referenced mulitple times in the de-duping scheme. Basically, I'm making the argument that if you're going to use de-duping- Raid-DP and WAFL is a MUST!
Regarding the caching- de-duping would eliminate the efficiencies of WAFL read-ahead, since the blocks wouldn't be temporally located. While the read ahead would pick up the stub, the de-duping software (not being netapp-aware) would need to go back and retrieve the actual data referenced by the stub, and there goes your cache hit %.
But, sometimes performance is not important.
Glenn (the other one)
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Walker Sent: Tuesday, February 27, 2007 8:41 AM To: Darren Soothill; toasters@mathworks.com Subject: RE: data management through compression for Netapp
I think that all of the solutions have their place - I'm not as interested in 'compression' as I am with 'deDupe' - they are separate, and not entirely equal.
The only issue I can foresee with deDupe, besides adding overhead, is chance of failure: if the data is stored in a single block, and it is lost, so are you - but that's why you have parity, so it shouldn't be any more dangerous than normal data storage (provided the software works adequately). As for overhead, having multiple individuals 'hit' the same block would cause performance issues - but I'm not so convinced: There's really no such thing as 'overwrite' in ONTAP, it's always writing to a free block - as for reading the single block that has been deDuped, even though multiple machines hitting the same block should ordinarily cause higher latencies for everyone, I would think that the simple fact that there is one block with many accesses would indicate that it would already be cached - caching it in memory should remove the read latency penalty.
Stubs carry their own penalties, btw: when the stub is seen, the filer must use the fpolicy service to communicate with the hosting system to redirect and\or obtain the actual data: this is multiple layers of network communication, memory lookup, and disk access before the client gets the requested data - this (in theory) should be slower than a self contained system in almost every case.
VTL is definitely a good solution for backups and we're looking into them now. But HSM products, and archiving products of all types, still have one common flaw: they move the data around and it must then be managed. While this would be the most likely best way of dealing with the massive amounts of data in anyone's environment, it would still be nice to keep it all in one place. That's just a dream, anyways... I'm not sure that there would be enough savings through deDupe to actually accomplish anything like this. Now on the other hand, if the filer were able to move data to cheaper ATA storage within the same system, and simply use a redirector in the filesystem (pointer through vvbn), that could prove useful: you'd have the same lookup of vvbn-to-pvbn in order to find the data, but the actual blocks would be elsewhere within the system. No additional layers of stubs, or redirection through some other HSM system.
Sorry for the long email this early in the morning,
Glenn
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Darren Soothill Sent: Tuesday, February 27, 2007 12:36 AM To: toasters@mathworks.com Subject: RE: data management through compression for Netapp
The issue with any compression technology is the latency it adds to any workload and what happens with a large number of people doing the same thing.
Would you not be better at looking at an archiving product that can archive off the older files leaving stub files behind so any application can carry on working with no interruption. Assuming that the data is stored as lots of files.
Perhaps you need to look at some traditional HSM type product that integrates with a tape library except instead of using a physical tape library you use a VTL which will do the deduplication and compression for you. You then have the added advantage that the VTL can spit out copies of the virtual tapes to be taken off site for DR. This type of solution should then be platform independent.
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Walker Sent: 27 February 2007 03:18 To: Glenn Dekhayser; Wilbur Castro; toasters@mathworks.com Subject: RE: data management through compression for Netapp
I've also heard good things about Data Domain.
On the other hand, does anyone have any experience with the A-SIS deDupe technology from NetApp? Sounds great, except for the 'out of band' nature of it.
Glenn
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Glenn Dekhayser Sent: Monday, February 26, 2007 8:36 PM To: Wilbur Castro; toasters@mathworks.com Subject: RE: data management through compression for Netapp
Wilbur:
What are your performance requirements? Adding compression/deDup into this has a serious impact on I/O performance, it will pretty much totally screw any read-ahead optimization.
If you still want to go ahead, look at Data domain's Gateway series. It doesn't formally support Netapp, but i'm sure you could use it. This will do exactly what you want, as long as you want NFS!
Glenn (the other one)
________________________________
From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com] On Behalf Of Wilbur Castro Sent: Monday, February 26, 2007 4:55 PM To: toasters@mathworks.com Subject: data management through compression for Netapp
Hi toasters,
We have a couple of 100 TBs of heterogenous storage (some from netapp), and see that our storage grows close to 60% over year. We were looking at alternatives for managing this data growth. Compression was one of the techniques we were considering for our nearline and (possibly) primary storage. Our applications cannot change to do their own compression, so it boils down to doing this in the storage layer or through an external device. Also, we'd like to not have any performance impact and compression to happen transparently. Deduplication technology from storage vendors would help, but it is not a hetrogenous solution.
I am not aware of any compression technology from netapp. Are you folks aware of any solutions? Would love to hear your experience with those or other alternative ways you deal with the storage growth problem while managing costs.
Thx, Wilbur
-
Glenn Dekhayser -
Glenn Walker