Present: James Gareth, Sam Govind Alessandra John Brian Chris Ewan, Pete Wahid Santanu Duncan 1. Put Ricardo on after Andrew. DPM - need roadmap, also WebDAV. Maybe a separate slot for WebDAV, or a slot for transfer protocols? WebDAV interoperation. WebDAV report: http://storage.esc.rl.ac.uk/grid-deployment/TR-Cirstea-2011.docx 2. James' report, and the following discussion: Most T1s considering splitting disk only away. Lots of people looking at filesystems such as OrangeFS (FNAL study) HTTP proxies and geolocation to "federate" dCache. Lots of people standardising their storage hardware, moving away from hardware RAID. RAID5 being a problem with large disk servers. RAIN with striping across nodes, or using replicating filesystems such as HDFS, things spanning across multiple nodes. Like Manchester did... Lustre - replicating things; grids also replicate things but are not so good at being fault tolerant. This is also the thinking about the stuff they call "federating" storage. The grid can't be totally failure tolerant, will rely on sites, because some data isn't yet replicated. Disk server on UPS gives some tolerance of power failure, Lustre has a feature like this as well. Advanced filesystems tend to delay writes, e.g. ext4. Linux XFS performs better on 3.0+ kernels, which we can't use because we are behind. XFS is most beneficial on the worker nodes; doesn't matter so much for the experiments code (but NFS4 support does). T1 used to compile custom kernels for the storage node. Useful to communicate to hepsysman Should we use ZFS? How well is it supported, and how is it licensed? [09:58:12] Govind Songara joined [09:59:56] Brian Davies joined [10:00:20] Alessandra Forti joined [10:00:32] John Bland joined [10:00:36] John Bland left [10:00:51] John Bland joined [10:00:52] Brian Davies left [10:01:50] Christopher Walker joined [10:02:04] Ewan Mac Mahon joined [10:03:11] Ewan Mac Mahon 1/2 hour sounds OK, but don't be too rigid about it. [10:03:48] Alessandra Forti hope evo is on [10:04:18] Wahid Bhimji joined [10:05:57] Alessandra Forti well it's the most used storage [10:07:47] Alessandra Forti not with in the house solutions. [10:08:13] Santanu Das joined [10:12:54] Sam Skipsey http://storage.esc.rl.ac.uk/grid-deployment/TR-Cirstea-2011.docx [10:13:01] Sam Skipsey (for those who don't want to find the email) [10:14:32] Wahid Bhimji you mean hepix [10:14:32] Ewan Mac Mahon Hepix? [10:14:36] Wahid Bhimji yes [10:15:27] Wahid Bhimji could you please talk a bit louder [10:16:15] Christopher Walker Moving away from hardware raid? [10:17:06] Wahid Bhimji EOS ; zfs ; RAIN - indeed [10:17:20] Alessandra Forti like WN? [10:17:30] Ewan Mac Mahon Is RAIN just an idea or something we have usable implementations of? [10:17:52] Duncan Rand joined [10:17:52] Duncan Rand left [10:18:28] Ewan Mac Mahon It's not completely clear to me what this would do to your reliability unless you're doing RAIN across nodes with RAID internally. [10:19:20] Ewan Mac Mahon And I'm not sure how easy it is to connect lots of drives to a machine with anything other than a RAID controller (you could run it in JBOD mode, but that saves you no money) [10:19:42] Ewan Mac Mahon RAI(grid sites) [10:20:00] Pete Gronbech joined [10:20:29] James Adams http://indico.cern.ch/getFile.py/access?contribId=38&sessionId=4&resId=0&materialId=slides&confId=160737 [10:20:54] James Adams if you are interested ewan [10:23:40] John Bland until someone pulls out the plugs from the server by mistake [10:24:00] Mark Norman joined [10:24:05] Ewan Mac Mahon Both the plugs [10:24:12] John Bland or the psu module fails [10:24:19] James Adams indeed [10:24:20] Ewan Mac Mahon Both the PSU modules [10:24:23] James Adams hence checksums [10:24:31] John Bland the psu housing can, and has failed, here [10:27:10] Jens Jensen It was the same problem with NFS4 - didn't work with the client using our old kernels [10:28:07] Wahid Bhimji way [10:28:25] Wahid Bhimji indeed - we'll be lucky to get sl6 - i have an sl6 pool node [10:28:46] Alessandra Forti i agree [10:28:59] Ewan Mac Mahon Well; I'm not sure that porting everyone to SL6 is necessarily easier than porting everyone to SL7; it's still a big porting effort. [10:30:06] John Bland another meeting, bye [10:30:09] John Bland left [10:30:49] James Adams remember that almost all the perfomance benefits from SL6/7 are from the kernel [10:32:19] Ewan Mac Mahon I think I'm currently favouring ext4 for new installs. [10:32:32] Ewan Mac Mahon But that's not based on anything particularly concrete. [10:32:34] Christopher Walker Lustre uses Ext4 underneath [10:32:44] James Adams XFS has the cooler name [10:32:59] Wahid Bhimji I can't give you the answer but I can give a little input on a couple of tests on SL6 box - indeed I don't have time to write an intro to all the FSs but I can show 2 plots [10:33:08] Sam Skipsey This is the "things with an X in them are awesome" argument, James? [10:33:13] Ewan Mac Mahon Ext4 has a '4'. Can't beat that. [10:33:14] James Adams yes [10:33:40] Ewan Mac Mahon and can I just say e-X-t4 [10:33:44] Wahid Bhimji no [10:33:52] Alessandra Forti X-t4 ? [10:33:58] Alessandra Forti [10:34:03] Ewan Mac Mahon [10:34:04] Wahid Bhimji indeed - not serious [10:34:25] James Adams it is allowed, in the same way as OpenAFS [10:34:30] Ewan Mac Mahon I think what I'd want to know about ZFS is that it's not stunningly better than the things I'm actually happy to run. [10:34:36] James Adams cannot be compiled INTO the kernel [10:34:39] Ewan Mac Mahon If it is, we might have a problem. [10:34:45] James Adams can be compiled FOR a kernel as a module [10:34:52] Ewan Mac Mahon If it's not, we can all get on with our lives. [10:35:16] Wahid Bhimji well lets have this discussion next week then [10:36:49] Alessandra Forti I've laready heard that one [10:39:41] Wahid Bhimji see you then [10:41:09] Brian Davies left [10:41:11] Wahid Bhimji left [10:41:14] James Adams left [10:41:16] Govind Songara left [10:41:16] Alessandra Forti left [10:41:17] Ewan Mac Mahon left [10:41:17] Santanu Das left [10:41:18] Mark Norman left [10:41:21] Christopher Walker left [10:41:25] Sam Skipsey left [10:41:27] Gareth Roy left [10:43:10] Pete Gronbech left