Ceph CDS Hammer, 29 Oct 2014

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: CDS Hammer (Day 2) - Shingled Erasure Code (SHEC)

Description

http://goo.gl/U4b70r

29 October 2014
Ceph Developer Summit: Hammer
Day 2
Shingled Erasure Code (SHEC)
Takeshi Miyamae, Takanori Nakao, Kensuke Shiozawa

A

B

All rights the next one we have here is the shingled erasure code with our friend from fujitsu, looks like our here. So if one of you, gentlemen, would like to take a few minutes to talk about what you have here and we can take it from there, hi.

C

Hi everyone, hello, my my my name is Takashimaya. My and the blueprint tighter is sheng gu, the laser code summary we have developed recovery efficient and I just average a call named shake we have we have presented. We have presented our our intermediate result at. I use Onix workshop ho depth 14 on october fifth, we'd like to provide the result for civil community, because we would like to get your feedback for our future. Work and owners are 44.

C

Uns 33 members, like l, takashimaya, my tagging, takanori nicole, and ken co-op members or owners, walk for 30 days limited and our current status is.

C

A

C

The working type 45 dry and it is currently being tested and waiting, and we think by frys user code, probing api and crush arisen, a subject for our purposes.

C

However, we have not started testing on giant or hammer yet and I will explain in detail. Detail description, check, throw buys more barriers, become very efficient. Layouts done deals. Hormone calls especially check is recovery, f, gent in case of multiple discs, firoz.

C

Figure figure 1 shows check, 10 6, 5, 6, 10, 6 5 has 10 data chunks and six RT chunks and the locality of the of each party chunk is 5 and.

C

Local locality, locality means each party's language or calculation ranges, for example, and in in the end of figure, 1 p, 1, p 1 is calculated from d12 d. 5 p-2 is calculated from d3 to d7, so we can. We can. We can confirm the route. The locality is 5, easy.

C

The by and by the way, for instance, if, if d, 6 and Denine break out check, we recover them with d3 and p4, because the union of calculate the the union of calculation ranges is smaller than the other unions of calculation ranges, for instance, p2, p3 and p4, p5 and so on. And in this case we must be.

C

We must read only six data or party chunks to recover d6 and t9.

C

The end so I'm in comparison with other.

C

Men other methods, the, for example, little sermon code.

C

Alma and our method is recover extent, because this sermon code has group global parties and.

C

We add an and it must it must read 1010 data or party chunks, but we we have, we. We only have to read six rich data or party chunks, so our method is a more recovery efficient and it.

A

Sorry at that point, quick question: in the case of the in the case of check 1065, then you're storing you have ten, you store 10 original data blocks and you store six additional parody blocks or shards. Yes, what's what's a, and the 5 number is how many, how many data blocks each shard or each parody covers that were the five is yes, yes,.

C

Ok, right Graham and necklace necklace, resected, specific specific. She recovery, efficiency and durability as usual called the properties and protect, reduce or mon coeur cold and shake rails on each three-dimensional property maps.

C

Figure 2 figure 2 is the property map of Lido Simone codes. In this case, upper area of the chart is sparse because little sermon code called doesn't have so many pick up very efficient layout.

C

On the other hand,.

C

If, in case of shake shake shake, I'll figure, 3 shows the the area the upper the upper area is filled, filled with check specific code cause. So, for example, the user who usually use rated 806 up our probe are provided I return, alternative.

C

Recovery efficient calls like check 652.

C

Therefore, we conclude that check is more adjustable down with someone caused by the way, in the next step, we'd like to extend checks, checks, layout and capabilities that could offer additional benefits to those of msn, russia and safe area. She msl she means Microsoft, arashi and Sierra. She means 6 p.m. ET kol. We are anticipating.

C

Anticipating the following benefits from those extensions: first, the extended chakra out include reed-solomon codes and ms energy, and they are all generated from a common parameter set. Second, the number of recovery option. Three outs is much larger down: nserror g or save energy and, finally, in case all safe crosta, that is, content content the in shingle dead center and which means that is not to split over one recovery. Overhead is more than sierra russia.

C

The this is the end of our presentation. Any questions I.

A

Assume there's um much to go look at the paper, but it sounds like the key advantage over. The current lrc approach is that you can deal with multiple, multiple disk failures still by still reading, only a subset of the data and Verity blocks is. That is that right, because currently with the with the current lrc, if you lose one disk, then you read just a subset of say: 515 or whatever. um But if you lose two, then you have to read all all 15 or 10 or whatever.

A

However many blocks to reconstruct, but here you only have to do two.

A

That is that right.

C

A

C

A

C

Duke smiles more so.

A

It sounds like the main advantage over f. Lr c is that you can repair a two-disc failure by reading. She is your skewer shards or discs or blocks, or whatever you call them is.

C

A

C

A

The is the recovery / ahead about the same or similar as far as the amount of additional data.

C

And income person walked so it was, she yeah m is Michael.

A

Shidoshi or Steph lrc I, don't I, don't haven't, read their work cell phone, so I'm not sure exactly how it works, but I'm guessing it's similar to what we did, but with with regular lrc, we restore say three more blocks and we get one third as many recovery.

A

So we trade we trade storage for recovery efficiency, but only with, but only for one disk. So here you're also you're, storing six additional blocks. I just thought, I think what I think I baby I just don't understand how to read the recovery efficiency on these plots to figure out cause. It.

C

Equal angle in comparison, I, inka, inka, inka, inka, solar in case of no more normal chak, chak de cabra efficiency and space efficient. She is what was down normal mmm in this era, russia, but, but we will. We like to extend our extend our way out and and the new new ray out has almost almost the same recovery efficiency and a specification I all well I made I made a mistake.

C

Happy may I may I explain from.

A

C

Headed in case of normal in case of normal shake, the the worst property is space efficiency and add to the durability.

C

C

The Dakota cover efficiency in case of multiple disk failures are good, but it is, it is rare, realized from the sacrifice of space, efficiency or durability.

C

C

C

So so we we like to extend our way out and.

C

We've dragged to a pub almost the same same ability with a meserole. She.

A

A

So this is, this is already you already implemented it with Firefly you're testing with the Firefly plugin interface, and it works yes, so I'm just there.

A

There are a couple: API changes with giant but I, don't think they're, big they're, pretty I, don't think they're hard to change adapt to, but I think we would be very interested in yeah adapting to the new plug-in interface and integrating it.

D

Yeah there was naught else change. It's not a big deal well, actually yeah strictly more powerful because of the lrc stuff. But if you don't pay attention to that, the it it's not important, but it's slightly more powerful and I think that's all yeah.

A

Okay, um yeah I mean I think it would be great to to merge this into the trees so that people can use it. The main I think the main thing that both are two things I think. First, you need to convert to the latest to the latest version of the API. Oh, the first thing would be to switch to the new version of the API and then the second thing I think there needs to be a like a written description.

A

That explains when you would want to use it and what advantages it has over either straight reed-solomon or the lrc or something else so that people know when when it makes sense and when it doesn't make sense and maybe what what parameters they would want to use, because if people are going to use it, they need to know why.

A

So, having said having a simple description of you know the cases where it's going to have big advantages and what those advantages are.

D

Hey just a quick question to make sure I understand what's going on um if you lose d1 and d2, is that recoverable.

C

A

Xd idea on the D 1 3 5 p-6.

D

Okie 1 and D 2, if you loo, if you lose those two is that recoverable.

C

Yes, we can begin acabo it. Oh yeah.

A

Prove he won p5 and p6 see why? Yes,.

C

D weekend this can recover from P 1 P 1, P, P, P 1, the p5 or p 0, p, 1, p, 6 or p5 and p6.

D

Well, I guess I misunderstood, I put it cool so.

A

I guess, as long as you have at least as many parity blocks, covering the range as you lost discs, then then your own, a cool okay, but within a range of five you can lose any given block is covered by three, but within any series of five you can use up to three discs: okay, I.

A

Think right, right, yeah, okay, you're it that's pretty cool.

C

Pretty cool, thank you. Thank you.

C

And may I ask some questions. Yes, yeah student and my first first question is the backward compatibility that does program interface. Obama have become possibility.

D

No, it's really really close. It's not gonna be hard to update it. We put it that way. This isn't more than an hour of work, including the time it takes to rebuild.

D

The primary difference is that we added the ability to reorder the data and parity chunks, for example, if you want to put parody one next to data 132 next, two day, 23 and parity, three next two day, 25 in the crush hierarchy, you can do that with the new version of the interface. That's the primary difference.

A

So I think I think what that what it means is that you would you would probably you would have a version for firefly and then.

E

You have a slightly.

A

Different version for a giant that just wires up to the slightly different interface, and we would have it in two different branches and it would be- and it's yeah it would be very easy.

D

Sorry about the line of stability we're just not quite it ready to commit to a nervous. Yet it's it's.

A

Brand new yeah the time that that ability to place Verity location.

E

D

I'm, sorry, oh yes, it is very important. It's it just doesn't yeah it doesn't make it hard to port. It just means that if you poured it and don't fill on that method, you get the old bear. That's all you can edit I'm. Sorry.

A

And this check could also take advantage of that placement, but community, oh yeah,.

D

Yeah apps absolutely.

E

D

That's why I that's the primary difference and they actually want it. So my guess is: it won't be that big deal for them.

A

This might be worth pointing out that there's one bit of work that we still need to do on the OSD, and that is um offloading recovery, work or whatever right.

D

So if you only need p for to recover well, let's say you will need p3 and p4 to recover d6 and d 9 or whatever you did. The primary is actually going to read all of the pieces locally and then write them back out. um It's not going to read any more than it needs to read. That's the good news. It's not going to rebuild the whole object, but it if the primary happens to be in the wrong row. It's still going to cause you a network hub, though the work we.

E

D

Is we need to burp the AC, a sequence of network transfers that perhaps you didn't want, though the work we had? The next thing is that we need to cause whichever replica is actually being recovered, to do the recovery locally or to do the reconstruction locally? It's not it's not a huge deal. It's just you know not.

A

Another thing to do yep and it's the same. It's the same thing that we need to do for LRC, for it to also take full advantage of the local, the local part of the local recovery right.

D

And that shouldn't require a change on your part when we asked the plugin. Actually it doesn't we'll just pick one of the recoveries yeah and tell that one to do it. Yeah.

D

Almost I'm wrong, I'm sure it'll cause a change. The energy I you and.

A

I, don't able at the audition tickets.

A

Yes, I think all the cost information is in the OSD another deciding.

E

A

Primary should be yeah right.

C

May I ask another question.

C

What what is our requirements for committing our source code into main trunk, for example? Do we need review process? Please predetermine the test process.

D

C

D

Will want to provide unit tests for it. It's a little weird you can talk to the week about how heated is the biggest thing is that we would like our art, our tautology tests, to include tests for this as well, that it will run in our regression sweet, um that's a little more involved, but not not hugely. So again, we can work with you on that, mainly though just follow the style guide, which shouldn't be terribly difficult,.

A

I'm. Writing the stand in the etherpad, but you can you can go look at that after.

A

I'm guessing for this, um what we probably ought to do another sub module. If the code for the the code for the Scheck is on github or something, then we could do another sub module if it's otherwise, if it's really small, then we can put it directly in the tree. But it's for.

D

This I'd rather do a subtree merge because it has backwards compatibility implications.

D

This isn't just a matter of any Eddie check will work that needs to be backwards, compatible or compatible with the last version of shack, so that I don't break people's culture. Is that kind of thing yeah.

A

But that's I mean the photo. Does that yeah.

D

Floor yeah, you don't tell people love them for this kind of thing I'd. Rather it were actually entry, yeah, okay. Firstly,.

A

Yeah either way um yeah so I type in the etherpad I wrote down a couple notes: I think it's you know moved to the probably move to the latest version of the API, although it doesn't matter that much write something at just some short document. That explains why you would use it so that users know when to use this and not something else, and some basic basic unit tests tests, hopefully um and then we'll integrate and then yes, I, met a pull request and we can review it like it. I get pull request.

A

Sounds great, oh ok,.

C

B

Cool looks like we got about 4 minutes left. If there are any questions prompt folks.

A

As as, has check been used in any other storage systems yet, or is this been pure research up until now,.

C

But I do under the end there there is a prong there.

E

So sorry for interruption, my name is Ken squishy. Without working with the NIA, my son, it is not run for using other storage storage appliances or anything. This point, cool and also the last question is that is there any do you have to decide who is going to commit to this check source code? Do are we responsible for maintaining, is at this point or is there anybody who who wish to join us.

D

Morning, I think you guys would be reviewing changes to it. Certainly, ok.

E

A

I mean, I think, I think if some random person showed up and wanted to change the check code, we would ask wait for you to okay it. Yes, we couldn't we're not going to be able to competently review or know whether that's a good change or not. Although.

B

I'm sure Luke will probably get his hands dirty at some point. Yeah.

A

Yeah I know but I mean hopefully others will write, object, that's kind of the idea, but yeah okay.

E

Look you're doing. Thank you. Yeah.

A

Awesome. Thank you all right. Great cool.