Re: [Gluster-devel] Glusterfs 3.1.2 dht/switch

[Anand Babu Periasamy]

Hi Gvozden,

We had an experimental capability in older releases called NUFA (Non Uniform Filesystem Access), exactly doing what you are asking. It allowed local disks to gain priority over remote disks.  NUFA required replication for high-availability. Replication in turn required network writes, there by limiting the benefits of NUFA.  In your case, data-availability is not a big deal. Unfortunately there was not much interest around this limited use-case feature, so we discontinued NUFA translator. There is currently no equivalent feature.

We are currently discussing internally about implementing HDFS capabilities to enable Map-Reduce jobs on top of Gluster. It is different from what you are asking though. Can you share more details about your application needs (type of data, IO pattern..). 

-AB

---

From: "Gvozden Neskovic" <address@hidden>
To: "Anand Babu Periasamy" <address@hidden>
Cc: address@hidden
Sent: Tuesday, March 8, 2011 6:09:06 AM
Subject: Re: [Gluster-devel] Glusterfs 3.1.2 dht/switch

Hi Anand,

Thank you for fast replay. 

I am looking into GlusterFS to setup cluster of about 600 nodes with ~2TB local disk

per node. This is only possible, with our workloads, if data locality is somehow enforced,

so that data 'produced and consumed' by node stays on that node's local disk, because there is no

dedicated network infrastructure. Data availability in case of node failures is not of big concern.

Do you have any suggestions regarding glusterfs setup I can use?

Best Regards,

-- 

Gvozden Neskovic
address@hidden

On Tue, Mar 8, 2011 at 1:57 PM, Anand Babu Periasamy <address@hidden> wrote:

Hi Gvozden, We have not tested switch translator in 3.1. Not enough interest around this translator as of now. My recommendation is to avoid this translator.

---

From: "Gvozden Neskovic" <address@hidden>
To: address@hidden
Sent: Tuesday, March 8, 2011 4:51:00 AM
Subject: [Gluster-devel] Glusterfs 3.1.2 dht/switch

Dear All,

I observed random misbehaving and crashes on setup with

switch xlator on 20 nodes.

Attached patch address following: 

memory leak in subvolume lookup, 

missing initialization in switch_create() and 

correct 'pattern.switch.case' option parsing. 

Best Regards.

 

-- 
Gvozden Neskovic
address@hidden

_______________________________________________
Gluster-devel mailing list
address@hidden
http://lists.nongnu.org/mailman/listinfo/gluster-devel