For map-reduce, the balancing is easier because you can configure slots. It would be nice to be able to express cores and memory separately, but slots are pretty good.
For HDFS, the situation is much worse because the balancing is based on percent fill. That leaves you with much less available space on smaller machines. You also wind up with odd segregation by age between different kinds of data. That leads to poor I/O performance. On Mon, May 2, 2011 at 10:31 AM, Jean-Daniel Cryans <jdcry...@apache.org>wrote: > I think the first issues you would encounter (just regarding HBase, > not talking about MR) is that if you have wildly different HW some > nodes might be able to handle their share of the load but some others > might not. At the moment the master doesn't know about the HW on the > slave nodes so it will just balance the regions equally. You would put > yourself in a situation where you would need to disable the balancer > and then do its job by yourself. > > Problems like that. > > J-D > > On Mon, May 2, 2011 at 10:03 AM, Chris Tarnas <c...@email.com> wrote: > > What are some of the common pitfalls of having different configurations > for different nodes? Is the problem more management issues, making sure each > type of node has its own config (so a 12 core box has 12 mappers and > reduces, an 8 core has 8, drive layouts, etc) or are there problems that > configuration changes can't deal with? > > > > thanks, > > -chris > > > > On May 2, 2011, at 6:57 AM, Michael Segel wrote: > > > >> > >> Hi, > >> > >> That's actually a really good question. > >> Unfortunately, the answer isn't really simple. > >> > >> You're going to need to estimate your growth and you're going to need to > estimate your configuration. > >> > >> Suppose I know that within 2 years, the amount of data that I want to > retain is going to be 1PB, with a 3x replication factor, I'll need at least > 3PB of disk. Assuming that I can fit 12x2TB drives in a node, I'll need > 125-150 machines. (There's some overhead for logging and OS) > >> > >> Now this doesn't mean that I'll need to buy all of the machines today > and build out the cluster. > >> It means that I will need to figure out my machine room, (rack space, > power, etc...) and also hardware configuration. > >> > >> You'll also need to plan out your hardware choices too. An example.. you > may want 10GBe on the switch but not at the data node. However you're going > to want to be able to expand your data nodes to be able to add 10GBe cards. > >> > >> The idea is that as I build out my cluster, all of the machines have the > same look and feel. So if you buy quad core CPUs and they are 2.2 GHz but 6 > months from now, you buy 2.6 GHz cpus, as long as they are 4 core cpus, your > cluster will look the same. > >> > >> The point is that when you lay out your cluster to start with, you'll > need to plan ahead and keep things similar. Also you'll need to make sure > your NameNode has enough memory... > >> > >> Having said that... Yahoo! has written a paper detailing MR2 (next > generation of map/reduce). As the M/R Job scheduler becomes more > intelligent about the types of jobs and types of hardware, the consistency > of hardware becomes less important. > >> > >> With respect to HBase, I suspect there to be a parallel evolution. > >> > >> As to building out and replacing your cluster... if this is a production > environment, you'll have to think about DR and building out a second > cluster. So the cost of replacing clusters should also be factored in when > you budget for hardware. > >> > >> Like I said, its not a simple answer and you have to approach each > instance separately and fine tune your cluster plans. > >> > >> HTH > >> > >> -Mike > >> > >> > >> ---------------------------------------- > >>> Date: Mon, 2 May 2011 09:53:05 +0300 > >>> From: iulia.zid...@1and1.ro > >>> To: user@hbase.apache.org > >>> CC: st...@duboce.net > >>> Subject: Re: Hardware configuration > >>> > >>> Thank you both. How would you estimate really big clusters, with > >>> hundreds of nodes? Requirements might change in time and replacing an > >>> entire cluster seems not the best solution... > >>> > >>> > >>> > >>> On 04/29/2011 07:08 PM, Stack wrote: > >>>> I agree with Michel Segel. Distributed computing is hard enough. > >>>> There is no need to add extra complexity. > >>>> > >>>> St.Ack > >>>> > >>>> On Fri, Apr 29, 2011 at 4:05 AM, Iulia Zidaru wrote: > >>>>> Hi, > >>>>> I'm wondering if having a cluster with different machines in terms of > CPU, > >>>>> RAM and disk space would be a big issue for HBase. For example, > machines > >>>>> with 12GBs RAM and machines with 48GBs. We suppose that we use them > at full > >>>>> capacity. What problems we might encounter if having this kind of > >>>>> configuration? > >>>>> Thank you, > >>>>> Iulia > >>>>> > >>>>> > >>> > >>> > >>> -- > >>> Iulia Zidaru > >>> Java Developer > >>> > >>> 1&1 Internet AG - Bucharest/Romania - Web Components Romania > >>> 18 Mircea Eliade St > >>> Sect 1, Bucharest > >>> RO Bucharest, 012015 > >>> iulia.zid...@1and1.ro > >>> 0040 31 223 9153 > >>> > >>> > >>> > >> > > > > >
