Upload
nicole-perkins
View
226
Download
0
Tags:
Embed Size (px)
Citation preview
Deployment metrics and planning
(aka “Potentially the most boring talk this week”)GridPP16
Jeremy [email protected]
27th June 2006
Overview
2 Even more metrics….zzZ
3 zzzzzzzzzZZZZZZZZZZZZZ
4 What came out of the recent deployment workshops
5 What is happening with SC4
6 Summary
1 An update on some of the high-level metrics
Available job slots have steadily
increased
Contribution to EGEE varies between 15% and 20%.From this plot stability looks like a problem!
0
1000
2000
3000
4000
5000
6000
06/2
0/04
07/1
1/20
04
08/0
1/20
04
08/2
2/04
09/1
2/20
04
10/0
3/20
04
10/2
4/04
11/1
4/04
12/0
5/20
04
12/2
6/04
01/1
6/05
02/0
6/20
05
02/2
7/05
03/2
0/05
04/1
0/20
05
05/0
1/20
05
05/2
2/05
06/1
2/20
05
07/0
3/20
05
07/2
4/05
08/1
4/05
09/0
4/20
05
09/2
5/05
10/1
6/05
11/0
6/20
05
11/2
7/05
18/1
2/20
05
08/0
1/20
06
29/0
1/20
06
19/0
2/20
06
12/0
3/20
06
02/0
4/20
06
23/0
4/20
06
15/0
5/20
06
05/0
6/20
06
Date
Pu
bli
shed
jo
b s
lots
UK total job slots
Thanks
to F
rase
r fo
r data
update
Our contribution to EGEE “work” done remains
significant but…
… but be aware that not all sites have published all data to APEL. Only 1 GridPP site is not currently publishing
CPU usage has been above 60% since May
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
06/0
2/20
04
06/2
5/04
07/1
8/04
08/1
0/20
04
09/0
2/20
04
09/2
5/04
10/1
8/04
11/1
0/20
04
12/0
3/20
04
12/2
6/04
01/1
8/05
02/1
0/20
05
03/0
5/20
05
03/2
8/05
04/2
0/05
05/1
3/05
06/0
5/20
05
06/2
8/05
07/2
1/05
08/1
3/05
09/0
5/20
05
09/2
8/05
10/2
1/05
11/1
3/05
06/1
2/20
05
29/1
2/20
05
21/0
1/20
06
13/0
2/20
06
08/0
3/20
06
31/0
3/20
06
23/0
4/20
06
17/0
5/20
06
09/0
6/20
06
Date
% j
ob
slo
ts u
sed
% EGEE slots used % UK slots used
Update for GridPP15
This is because most VOs have doubled job rates –
note LHCb!
IC-HEP are developing a tool to show job histories (per CE or Tier-2)
View for GridPP CEs covering last week
..but it looks a little rough sometimes!
Over 5000 jobs running
The largest GridPP users by VO for last 3 months
LHCb
ATLAS
BABAR
CMS
BIOMED
DZERO
ZEUS
VOs = a big success
0
5
10
15
20
25
Nu
mb
er o
f en
able
d V
Os
Jan-06
Jun-06
• But we do now need to make sure that schedulers are giving the correct priority to LHC VO jobs!• The ops VO will be used for monitoring from the start of July
Ranked CEs for Apr-Jun 2006
Thanks
to G
idon a
nd O
livie
r fo
r th
is p
lot.
Ranked CEs for Apr-June 06
Successful time / total time
Thanks
to G
idon a
nd O
livie
r fo
r th
is p
lot.
An interesting view by Tier
0
100000
200000
300000
400000
500000
600000
700000
800000
London Tier-2 NorthGrid ScotGrid SouthGrid RAL Tier-1
To
tal H
ou
rs
Failed Hours
Success Hours
A little out of date Q1 view for contribution and occupancy
0.00%
20.00%
40.00%
60.00%
80.00%
100.00%
120.00%
Average occupancy Contribution to UK Tier-2 processing
Some sites appear more successful at staying full even when overall job throughput is not saturating the resources. For Q2 most sites should show decent utilisation. (of course this plot involves estimates and assumes 100% availability).
Storage has seen a healthy increase – but
usage ~40%
SRM V2.2 is delayed. There have been several workshops/meetings taking forward the details of storage types (custodial vs permanent
etc.)
Scheduled downtime is better than EGEE average
…. Still not really good enough to meet MoU targets. Sites need to update without draining site… there are still open questions in the area of what “available” means. GOCDB needs finer granularity for different services.
So are there any recent trends!?
0
20
40
60
80
100
120
140
160
UCL-Cen
tral
Brunel
Sheffie
ld
IC L
eSC*
Cambridge
Queen
Mar
y, UL
UCL-HEP
Royal H
olloway
, UL
Durham
Liver
pool
Edinbu
rgh
IC H
EP
Glasgo
w
RALPP
Birming
ham
RAL-LC
G2
Bristo
l
Oxford
Lanc
aster
Man
ches
ter
Sta
cked
% s
ched
ule
d d
ow
n e
ach
mo
nth
May
April
March
February
January
This is the percentage of time that a site was down for a given period – if down for whole month the monthly stack (each colour) would be 100%
% SFTs failed for UKI
Seems better than the EGEE
average for April and May but
slightly worse in June so far.
These figures really need translating into hours unavailable and the impact on the 95% annual availability target.
SFTs per site - time
0
2
4
6
8
10
12
14
RAL-LC
G2
IC L
eSC*
RALPP
Glasgo
w
IC H
EP
Queen
Mar
y, UL
UCL-HEP
Durham
Man
ches
ter
Lanc
aster
Liver
pool
Bristo
l
Oxford
Birming
ham
Cambridge
Edinbu
rgh
Brunel
Sheffie
ld
Royal H
olloway
, UL
UCL-Cen
tral
Sta
cked
% o
f S
FT
s fa
iled
eac
h m
on
th
May
April
March
February
January
0
10
20
30
40
50
60
70
80
90
Queen
Mar
y, UL
Durham
Sheffie
ld
UCL-HEP
Man
ches
ter
Oxford
IC H
EP
RAL-LC
G2
Glasgo
w
Brunel
RALPP
IC L
eSC*
UCL-CENTRAL
Edinbu
rgh
Liver
pool
Cambridge
Lanc
aster
Birming
ham
Royal H
olloway
, UL
Bristo
l
Sta
cked
%s
of
tim
e u
nav
aila
ble
eac
h m
on
th (
SF
T)
May
April
March
February
January
Generally April and May seem to be improvements on January to March
Number of trouble tickets
0
2
4
6
8
10
12
14
16
18
RAL-LC
G2
QMUL-e
Science
Man
ches
ter
Durham
IC-L
eSC
Oxford
Lanc
aster
IC-H
EP
Birming
ham
Brunel
Liver
pool
Edinbu
rgh
RHUL
UCL-Cen
tral
RALPP
Cambridge
Glasgo
w
UCL-HEP
Sheffie
ld
Bristo
l
Nu
mb
er o
f n
ew t
icke
ts
Q1-2006 Q2-2006
More tickets in Q2 2006 so far! This seems correlated with the increased job loads. The profile is really quite similar between Q1 and Q2 2006
Average time to close tickets
0
20
40
60
80
100
120
140
160
RAL-LC
G2
QMUL-e
Science
Man
ches
ter
Durham
IC-L
eSC
Oxford
Lanc
aster
IC-H
EP
Birming
ham
Brunel
Liver
pool
Edinbu
rgh
RHUL
UCL-Cen
tral
RALPP
Cambridge
Glasgo
w
UCL-HEP
Sheffie
ld
Bristo
l
Ave
rag
e ti
me
to c
lose
tic
kets
(h
rs)
Q1-2006 Q2-2006
Tickets are usually from grid operator on duty. We need to look at factors behind these times. Note that just a few tickets staying open for a long time can distort the conclusions. We need better defined targets. The MoU talks about time to response of 12hrs (prime time) and 72 hrs (not prime time).
Middleware upgrade profiles remain similar
0
5
10
15
20
25
30
35
40
09/0
4/20
05
23/0
4/20
05
07/0
5/20
05
21/0
5/20
05
04/0
6/20
05
18/0
6/20
05
02/0
7/20
05
16/0
7/20
05
30/0
7/20
05
13/0
8/20
05
27/0
8/20
05
10/0
9/20
05
24/0
9/20
05
08/1
0/20
05
22/1
0/20
05
05/1
1/20
05
19/1
1/20
05
03/1
2/20
05
17/1
2/20
05
31/1
2/20
05
14/0
1/20
06
28/0
1/20
06
11/0
2/20
06
25/0
2/20
06
11/0
3/20
06
25/0
3/20
06
08/0
4/20
06
22/0
4/20
06
06/0
5/20
06
20/0
5/20
06
03/0
6/20
06
17/0
6/20
06
# si
tes
at r
elea
se
LCG-2_6_0 LCG-2_7_0 GLITE-3_0_0 LCG-2_4_0
• gLite 3.0.0 was deployed late but released on time raising questions about project wide communications. Our target remains 1 month from agreed start date.
• EGEE wants to move to “rolling updates” but there are still issues around tracking (publishing) component versions installed.
Disk to disk transfer rates
0
100
200
300
400
500
600
700
800
900
Lanc
aster
RALPP
Birming
ham
Glasgo
w
Edinbu
rgh
Oxford
Man
ches
ter
Sheffie
ld
Cambridge
Bristo
l
UCL-CENTRAL
Durham
QMUL
IC-H
EP
IC-L
eSC
UCL-HEP
RHUL
Brunel
Liver
pool
Tra
nsf
er r
ate
in M
b/s
Inbound Outbound
• The testing went well (thanks to Graeme) but we have a lot to do to improve rates.• Suspected/actual problems and possible solutions are listed in the SC wiki:http://www.gridpp.ac.uk/wiki/Service_Challenge_Transfer_Test_Summary
Some key work areas for Q3 and Q4 2006
• Improving site availability/monitoring (e.g. Nagios scripts with alarms)• Getting the transfer rates higher• Understanding external connectivity data transfer needs• Understand performance differences across the sites• Adapt to rolling update of middleware model• Implement storage accounting• Improve cross-site support• Understand WLCG MoU mapping to UK Tier-2 structure (and how we meet it)• Take part in LCG experiment challenges (SC4 and beyond)• Streamlining of the support structure (helpdesk)• SRM upgrades (SRM v2.2)• New resources integration (start to address the CPU:disk imbalance vs requirements)• Security: incident response • Exploiting SuperJanet upgrades• Improved alignment with UK National Grid Service• The usual: documentation and communication
Workshop outputs
Tier-2 workshop/tutorials already covered – next planned for January 2007
OSG/EGEE operations workshop
RELEASE AND DEPLOYMENT PROCESS– Why do sites need to schedule downtime for upgrades?– Release: Is local certification needed? sites required for testing against batch systems– Links to deployment timetable and progress area
USER SUPPORT– How to improve communications (role of TCG was even debated!)– Experiment/VO experience. Improving error messaging!
SITE VALIDATION– Site Availability Monitoring (SFTs for critical services – will remove some of the general SFT
problems that end up logged against sites)
VULNERABILITY & RISK ANALYSIS– New in EGEE-II = SA3. – Move to a new policy for going public with vulnerabilities – RATS (risk analysis teams)
Service Challenge technical workshop– Review of individual Tier-1 rates and problems– Experiments plans are getting clearer and were reviewed– Commitment to use GGUS for problem tickets
Identified experiment interactions (please give
feedback!)
ScotGrid (Signed up to ATLAS SC4)DurhamEdinburghGlasgow – PPS site involved with work for ATLAS
NorthGrid (Signed up to ATLAS SC4)Lancaster – Involved with ATLAS SC4LiverpoolManchester – Already working with ATLAS but not SC4 specific Sheffield
SouthGridBirminghamBristolCambridgeOxford – ATLAS?RAL-PPD – Will get involved with CMS
London Tier-2Brunel – Offer to contribute to ATLAS MC production.Imperial – Working with CMSQMUL – ATLAS? (manpower concerns)RHUL – Bandwidth concern. ATLAS MC?UCL
Summary
2 Within EGEE and WLCG our contribution remains strong
3 Some issues with SFTs and scheduled downtime
4 Workshops over last 2 weeks have been useful
6 We need more sites to be involved with experiment challenges
1 There is a lot of data but not in a consistent format
5 Some clear tasks for next 6 months