Attending
	Brian, Elena, John H, Matt, Ewan, Jeremy, Sam, Robert, John B, Gareth,
	David, Chris W, Rob, Wahid, Chris B, Govind, Jens


1. Hardware recommendations - discussion

How to do the discussion - share experiences, usually whoever went through
procurement most recently (probably T1).

Ewan bought servers for his "other cluster', R720XD, 12 bay, which is "the new
R510"  Lancaster talked to Viglen about 24 or 36 bay servers.

Some discussion as to whether system disks should be on separate drives
(Glasgow) or not (Oxford).  E.g. with dodgy RAID cards, would make the OS
available if not on the RAID system - but you might have bigger problems with
dodgy RAID.  Another option is to put the OS on a memory stick - could be
explored further.

A press release announced "6TB drives 'soon'" (next year), which might make
4TBs cheaper.  Probably now roughly the same cost/TB between 3TB and 4TB, but
with larger drives also consider the rebuild time.  And the rate of
reading/writing to the drives.

Ewan looked at 60 bay 4U - less space, less energy, more storage - compared to
smaller, separate units - but then if the thing goes down, everything is going
down - CASTOR took the approach of having >= five boxes in a service class, so
no more than 20% is lost if one goes down.  For the 60 bay thing, also need to
consider that a RAID set with our current cards can have at most 32 drives.

Chris B looked at WD SE series (rather than RE).  Measuring lifetime in TB/yr
- ? - but see also presentation (link in chat) from HEPiX about failure modes
and workload dependency.

T1 will always get diversity, so as to not lose everything if there are
problems with eg firmware or something from a given vendor (seen before.)

Only QM, Glasgow, and Lancaster are *not* doing procurement this time around.
Need to get going ASAP if the lot is to be completed before the end of the FY.
Also, some vendors are quicker to supply than others.  Specify delivery time
in the ITT.

4. AOB

Empty directories - see chat.

[10:14:17] Ewan Mac Mahon The benefit I care about is the cheaper purchase price.
[10:14:31] Ewan Mac Mahon There may be other things, but that's what I care about in this case.
[10:14:37] Wahid Bhimji are you talking about getting a big dense storage bo
[10:14:41] Wahid Bhimji box
[10:14:56] Ewan Mac Mahon We're musing about it.
[10:14:57] Jens Jensen Yep
[10:15:19] Wahid Bhimji but then in your Oxfords specifc case you will have an interesting mix if your current is R510s
[10:15:54] Wahid Bhimji not that that s a problem but its a swing
[10:16:07] Ewan Mac Mahon Indeed. It's an interesting mix already, some 16 bay Supermicros, a few 24s, three 36s, and a whole bunch of the R510 12x 3TB servers/
[10:17:05] Ewan Mac Mahon Indeed, our 36 bays are 2x 18 drive arrays, but that's on a single controller/
[10:17:15] Ewan Mac Mahon And a single network interface etc.
[10:17:48] Matt Doidge I wonder if the 36 bay would be better on 2 controllers, or would that not fit/be too costly
[10:18:19] Wahid Bhimji Have you talked about DDN
[10:18:23] Ewan Mac Mahon lol
[10:18:32] Wahid Bhimji are you laughing at me
[10:20:09] Jens Jensen There's a DDN meeting tomorrow somewhere in London. Shaun is going.
[10:20:27] Sam Skipsey We bought 36bay units for both cost efficiency *and* space efficiency reasons, actually.
[10:20:32] Jens Jensen It's more irods related (ie WOS), I believe
[10:23:19] Ewan Mac Mahon Indeed, the 12 bay ones are not terribly high density
[10:25:04] Ewan Mac Mahon The thought has crossed my mind.
[10:25:19] Ewan Mac Mahon I might just run cloud instances on them though.
[10:25:38] Sam Skipsey (We *did* try putting some storage as batch nodes. It turned out the WN and DPM-disk packages were incompatible. Might work better as VMs or Containers, which we have thought about.)
[10:25:57] Ewan Mac Mahon Do the PMB think about things to that sort of level?
[10:26:44] Matt Doidge Sam - use the WN tarball?  
[10:27:07] Sam Skipsey We also thought about that. Tarballs in Containers might be nice  
[10:27:08] Ewan Mac Mahon We did look a while ago at building a DIY box out of consumer grade parts and it wasn't a hell of a lot cheaper.
[10:28:12] Jens Jensen Can't be much different per TB. But do you get the same bandwidth
[10:28:34] Wahid Bhimji yeah thats what I am worried about - another performance drop
[10:28:41] Jeremy Coles In answer to Ewan - no. The PMB is not giving direction on the technology. Those decisions come out of discussions like the one today.
[10:28:47] Wahid Bhimji IO per tb
[10:29:29] Matt Doidge That would affect servers with fewer disk slots more.
[10:29:59] Ewan Mac Mahon @Jeremy It wasn't a technology question as such, it was a matter of how much funding is issued to get a given amount of hardware.
[10:30:27] Matt Doidge Does anyone know the HEPSPEC/TB ratio we should be aiming for?
[10:30:47] Ewan Mac Mahon AIUI, in practice, the two figures are mostly arrived at independently, we get handed an amount of money, and we get handed a requirement. The one does not greatly inform the other.
[10:31:58] Jens Jensen We'll get someone from the T1 fabric team next week (they're busy todya with the power stuff)
[10:32:41] Matt Doidge At certain points in the past we've been given ratios to aim for. But now I get the feeling the VOs themselves have no idea!
[10:33:27] Jens Jensen This one? http://indico.cern.ch/contributionDisplay.py?contribId=37&sessionId=3&confId=247864
[10:33:42] Ewan Mac Mahon On Supermicro vendors; Oxford's central purchasing folks seem to have positive feelings about OCF at the moment, so they're on our potential vendors list along with the usual viglen.
[10:33:44] Wahid Bhimji Thanks
[10:34:25] Ewan Mac Mahon I'm planning to mostly ignore ratios and buy 'enough' disk and as much CPU as possible.
[10:34:57] John Bland define "enough"
[10:35:12] Ewan Mac Mahon Well, hmm.
[10:35:22] Jeremy Coles Ewan: To a first level that is correct. The money is 'fixed' and gets divided in proportion to the experiment shares and this then has to be used to meet our WLCG pledges. There is an indicative requirement to guide the site wrt capacity.
[10:35:32] Chris Brew I was half thinking about trying to work out whther a pound spent on storage or CPU brought in a greater return
[10:35:43] Ewan Mac Mahon I want to stay comfortably but not excessively over pledge, but there's quite a bit of guesswork involved in that.
[10:36:03] Matt Doidge Especially as our kit ages.
[10:36:04] John Bland does anyone know what the pledges will be/are yet?
[10:37:24] Jeremy Coles Pete Gronbech has estimates. I think Dave Britton is checking them before circulating, perhaps later this week.
[10:39:29] Matt Doidge At Lancaster we'd often just use the NSSA, which keeps tendering times down.
[10:39:58] Sam Skipsey I did initially read that as the "NSA", Matt, which did make me wonder what kind of deals you guys have been doing...
[10:40:48] Matt Doidge Nothing like blackmail and dirty tricks to ensure swift delivery.
[10:43:06] Ewan Mac Mahon Well, it does mean that there's enough time for Glasgow and QMUL to have these discussions with DDN.
[10:43:33] Ewan Mac Mahon Aside from the pricing, I just don't think we've got time to consider doing anything radically new and different.
[10:43:41] Wahid Bhimji  
[10:47:12] Chris Brew find /pnfs -type d -empty -print -delete
[10:47:38] Ewan Mac Mahon I think Chris just did  
[10:47:48] Ewan Mac Mahon ^^^^
[10:49:42] Wahid Bhimji thanks bye