Attending:
	Gareth, Govind, John H, Matt, Wahid, Brian, Robert, Jens, Sam, Pete,
	Elena, Ewan, Raul, Chris W, David, Steve

1. Feedback on Wahid's slides for DPM workshop tomorrow (sent to list 10.12.13)
   (at 14:15?)

Milestone reached - all DPM sites on 1.8.7!

CASTOR publishing multiple versions of stuff.  In the future DPM won't have "a
version number" either; more like versions of things - and experiments'
requirements may also be "need DAV at blah" rather than having the whole lot
at a particular version.  Nevertheless, the bits all need to work together, so
some sort of combined baseline and maybe a "safe" combination would be needed.

We mentioned S3; would it be worth picking up S3 again?

There's a DDN writeup from Sam, a link to which has been sent to Wahid, it has
plots in it (so it's a proper writeup).

Brian's other point from earlier about limiting the number of connections is
also worth mentioning.  Limiting the number of threads does not limit the
number of connections, just slows stuff down.

There is some xroot throttling stuff from Brian Bockelman - what's the status
of that?

LHCb getting xroot TURLs at Manchester; problem is understood and a ticket is
open and it is progressing.

It is worth mentioning that the support that worried us so much last year
seems to be working - there is a community support, and it is good to see
other members of the community attending the meeting (people from Italy,
France, CzR.)  But above all the DPM guys from CERN should be lauded for the
quality and consistency for their support.

Wahid noted that we are not getting experiment input to this talk (or the
workshop in general).

As an aside, maybe Google+ posts and Tweets should count as dissemination as
well - currently the metric is the blog posts - which we now consistently fail
- maybe the other stuff should count, too.  Probably Tom has the key to the
Twitter account.

2. Hardware discussion

Different mfrs use different disks - in Oxford the recent winner used 4TB
disks (as were the T1 ones).  R720 XD, not using MD1200.  Quote specified SATA
drives.  Wahid has 1-2 MD1200 behind each server.

Planning as a guideline 1GB RAM per 1 TB of usable storage.

3. Feedback on small VO policy, particularly wrt space tokens

Small VOs may have the unfortunate habit of growing and become large VOs (as
in "all VOs start out small one day" but that is probably not true.)  What are
the incentives for them using space token - if sharing space, they can
use up the common space , just like a VO member could use up the space
allocated to a single VO: probably this is just something that has to be
sorted out with the LART (by the VO).

DPM can add more space to a space token; dCache (probably) treats it slightly
differently, ie less as an abstraction.

Brian asks the question about whether to put them into D1T0 or D0T0.  Probably
users would not like if their file was GC'ed without their knowledge, but
there may be a use for this.

Perhaps it should be a requirement for users to talk to us more often... like
the T1 experiments liaison meeting which is weekly, but we could do something
less frequent but still regularly occuring.

Exercise is useful though, to update the documentation for the small VOs.

Also, of course, when small VOs use non-SRM tools, how does the space token
management work...?

4. AOB

Jens is at RHUL next Wed.,; will dial in if schedule is OK, otherwise not.
In case of not, Jens thanked everyone and wished everyone merry.


[10:02:08] Jens Jensen http://indico.cern.ch/conferenceDisplay.py?confId=273864
[10:09:40] Matt Doidge The grass is always greener on the other side (of the SRM).
[10:13:40] Pete Gronbech Are we meant to be able to see your slides at the mo??
[10:13:54] Jens Jensen I thought it was a bit futher ahead...
[10:14:10] Jens Jensen @Pete: the slides are in a mail sent to the list
[10:14:28] Pete Gronbech ok got it
[10:16:32] Matt Doidge Lancaster's Pet Peeve is Hotspotting.
[10:18:00] Elena Korolkova Where are actual nimbers are coming from at slide1?
[10:18:37] Elena Korolkova According to gstat and rebus we have 360 TB in Sheffield
[10:18:52] Ewan Mac Mahon Do we want to note the absence of IPv6 xrootd support? I know that's arguably not reallly a DPM thing, but it's a thing.
[10:26:58] Christopher Walker Can you point me at the talk...
[10:27:14] Jens Jensen Chris, it was sent to the list yesterday
[10:31:21] Ewan Mac Mahon Does this tick any metrics?
[10:31:40] Ewan Mac Mahon I'm sure Wahid will be live-blogging the whole thing, which should help.
[10:32:27] Ewan Mac Mahon Google+ posts count as publications, right?
[10:33:31] Sam Skipsey Obviously we need a tumblr...
[10:35:34] Matt Doidge I'm fine to chat now.
[10:36:05] Matt Doidge Lancaster got quoted by Viglen 4TB that weren't quite as cost effective as 3TB
[10:36:28] Ewan Mac Mahon That's wierd. Viglen quoted us 3TB ones.
[10:37:41] Elena Korolkova I've got a quote from Dell with 4TB disks
[10:38:28] Matt Doidge I asked specifically for a 4TB quote. It works out a little cheaper per TB for 24 bay servers, but mroe expensive per TB for 36 bay servers...
[10:41:14] Elena Korolkova What do people think about PowerVault MD1200?
[10:41:51] Matt Doidge What's the RAM per TB ration people have been going for?
[10:42:09] Ewan Mac Mahon @Elena I think we think they're fine, but I'd rather just have the servers.
[10:42:37] Matt Doidge *ratio, although I suppose ration works.
[10:43:16] Matt Doidge 32GB of RAM is peanuts to put into a server, 64GB is still only a few hundred quid.
[10:44:02] Wahid Bhimji yeah 64 is what I got specced but if I put 2 MD1200s behind my R720xd I might try and double it
[10:47:45] Ewan Mac Mahon I think that's us, isn't it?
[10:48:03] Ewan Mac Mahon (the sentient feedback loop)
[10:48:30] Matt Doidge Can anyone remind me what the (supposed) "optimum" number of disks in a raid 6 is please ? I think the formula went something like 2N+X.
[10:50:08] Ewan Mac Mahon X=2  
[10:50:24] Ewan Mac Mahon And yes, I think that's right.
[10:50:40] Ewan Mac Mahon Which makes the 12 drive machines work out quite nicely.
[10:51:06] Matt Doidge Ta!
[10:51:12] Pete Gronbech Exactly, at Oxford we created a 2tb spacetoken for t2k but due to some issue ewan can remind me about they are not using it but have ~30TB in the unallocated area.
[10:51:35] Sam Skipsey (It's also tricky for dCache sites, IIRC how their spacetoken allocation works, since being dynamic is harder for them.)
[10:52:00] Pete Gronbech actually 37TB!! which is more than I would like
[10:52:06] Steve Jones Maybe "unallocated area" should be very slow, or some other cost of using it?
[10:52:22] Wahid Bhimji @ewan / matt unless you want a hotspare (on a 12 drive its a bit of a waste but on the 36 slots less so (but then can't do 2N + X)
[10:52:55] Steve Jones We need a stick as well as a carrot. Charge to use "unallocated area", i.e. more job slots if you use a token?
[10:53:03] Ewan Mac Mahon Actually, is that formula right?
[10:53:19] Ewan Mac Mahon It sounds familiar, but doesn't that basically reduce to 'even number of disks'?
[10:53:20] Matt Doidge Surely small VOs would *want* a space token, as it 's guranteed space.
[10:53:37] Steve Jones We need a scheduler that penalises jobs that use "unallocated area" and rewards jobs that use "space token".
[10:53:54] Sam Skipsey Matt: sure, but it looks more complex to them.
[10:54:05] Ewan Mac Mahon The downside for the sites is that if you're supporting 20 VOs and give each of them a 2TB spacetoken, that's a 40TB server burnt on day 1, with no actual data.
[10:54:10] Sam Skipsey Plus, if only N of M supporting sites require one...
[10:54:46] Sam Skipsey then it means VO X has to do something Different when storing on those sites compared to the (M-N) of them.
[10:55:10] Sam Skipsey (And this can't be fixed just by us doing the same thing in the UK, as some of these Medium Sized VOs are International)
[10:56:06] Steve Jones If space token use was transparent, then it would not look more complex. Idea: use implcit space token, i.e. assumed default.
[10:56:25] Sam Skipsey Sure, the ideal would be to allow an SE to have "VO quotas"
[10:56:52] Steve Jones How hard to implement?
[10:57:17] Steve Jones Yes, we need proposals.
[10:57:27] Steve Jones Feedback = routines.
[10:57:44] Ewan Mac Mahon Quotas are different to space tokens on purpose though - the advantage for the VOs is the reservation aspect. The advantage for the site is the quota aspect.
[10:57:53] Ewan Mac Mahon You sort-of do want both.
[10:58:29] Sam Skipsey Indeed
[10:58:38] Ewan Mac Mahon Also, that's part of the answer to encouraging people to use them - the promise is that if they use a space token, the space will be there when they want it, if they don't, it might not.
[10:58:43] Sam Skipsey And SRM was never designed to support Quotas  
[10:59:30] Wahid Bhimji haha
[10:59:33] Steve Jones Cheers!
[10:59:37] Wahid Bhimji ok - I'll be here next week