EMC CUSTOMER UPDATE · PDF file12 juni 2012 – Fort Voordorp ... HBA drivers ----- SAN...
Transcript of EMC CUSTOMER UPDATE · PDF file12 juni 2012 – Fort Voordorp ... HBA drivers ----- SAN...
1 © Copyright 2012 EMC Corporation. All rights reserved.
EMC CUSTOMER UPDATE 12 juni 2012 – Fort Voordorp
What is Performance? Dependencies in Performance? What is expected? Jan Sterk en Jeroen Kleinnibbelink
2 © Copyright 2012 EMC Corporation. All rights reserved.
Performance Impact Hierarchy
Application Code -----------------------------------------------------------------------------------
Database Layer -------------------------------------------------------------------------------
File System ----------------------------------------------------------------------------
Logical Volume Manager -------------------------------------------------------------------------
OS Kernel ----------------------------------------------------------------------
HBA drivers -------------------------------------------------------------------
SAN ---------------------------------------------------------------
Storage
Request
3 © Copyright 2012 EMC Corporation. All rights reserved.
Virtualisation impact
Request
VM/LPAR ------------------------------------------------------
ESX/VIOS ----------------------------------------------------
VPLEX/SVC ------------------------------------------------
SAN --------------------------------------------
Storage
Measure ------------------------------------------------------
Measure ---------------------------------------------------
Measure ------------------------------------------------
Measure --------------------------------------------
Measure
5 © Copyright 2012 EMC Corporation. All rights reserved.
(Sublun) Skew
Skew is the distribution of I/Os over a defined capacity. The concept can be associated to the 80/20 rule or the "Pareto law".
In this sample, the Workload Skew is 88% / 12%, indicating that 82% of I/Os executed are in 12% of the used capacity.
10 © Copyright 2012 EMC Corporation. All rights reserved.
Storage Pool Multiple Thin Devices
Allocation units (768 KB = 12 tracks)
“TDAT”
“extents” or “chunks”
“TDEV”
Thin Terminology VMAX
Enabling technologies: 1.TDEV,TDAT, and Pools 2.IVTOC
11 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP VMAX 5875/5876 FAST VP tunes application I/O with sub-LUN moves (minimal 6
extents) Improve performance with fine-grained data placement and VP
wide striping – Reduce drive contention
Adapt to workload and policy changes Reduce TCO with precise FLASHSATA Capacity Balancing Push more data to economical SATA drives Maximize the benefits of FLASH Drives
– Only move the high performance data to FLASH Drives
Optimize performance in Oracle, SQL, SAP and VMware environments
Virtual LUN VP Mobility provides full user control for re-tiering
12 © Copyright 2012 EMC Corporation. All rights reserved.
Initial Performance Analysis DMX-3 SAP Analysis – August 2010
Frame DMX 128 GB Cache 32 Front-end ports 16 Disk channels 248 146 GB10 K 263 300 GB 15 K 90 300GB 15 K Enginuity 5773
Max throughput for
batch
13 © Copyright 2012 EMC Corporation. All rights reserved.
Post Migration Analysis – Pre FAST-VP March 19, 2011
Frame VMAX-2 128GB Cache 32 Front end ports 17 400GB EFD Drives 314 450GB FC Drives 90 2TB SATA
30% Less
Drives
• Initial Post Migration workload to the VMAX
• SATA drives are hot due to data migrating all of the SAP Test/Dev environment to the SATA drive pool − Initial throughput is at 1 GB/s for Batch − Initial IOPS is approximately 30,000
IOPS − SATA drives are hot as all Test/Dev is put
on the slow pool with NO FAST policy − Initially the EFD drives are idle
14 © Copyright 2012 EMC Corporation. All rights reserved.
Post Migration Analysis – FAST VP Initiated March 23, 2011
• Initial Post Migration workload with first FAST VP Policy applied to VMAX • SATA Drives are hot due to their conservative policy for the EFD drives
15 © Copyright 2012 EMC Corporation. All rights reserved.
Post Migration Analysis – FAST VP Policy Adjusted April 7, 2011
2x Through
put
• Workload with adjusted FAST-VP Policy applied to the VMAX.
• SATA Drives are cooling off and the EFD drives are starting to get more workload applied.
• More importantly, the workload doubled!
• Virtual LUN Migrator is used to do ‘on the fly’ migrations of hot tables.
17 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP Principles • Pools are the new paradigm for storage
– An association of drives to be shared among many workloads • Pool LUNs can be Thin or Fully Provisioned
– ‘Thin’ for space efficiency – ‘Thick’ for performance bias
• Pools are easy to set up, and distribute – Any kind of disks, in any (back-end) location – Thin and thick LUNs can share a pool – Pool LUNs can be EXPANDED – Pool LUNs can be SHRUNK (windows VDS) – Pool LUNs can be COMPRESSED
• Pools can be made of multiple tiers – SSD, SAS, NL-SAS
18 © Copyright 2012 EMC Corporation. All rights reserved.
Pools and Allocation • Thin LUNs allocate 3 GB at creation
– Over 1 GB is metadata; allocates more 1 GB slices as needed – Allocates subunits within 1 GB slice as small as 8 KB: “packs together” sparse
allocation – Best for lower performance, high TCO
• Thick LUNs allocate 2 GB, reserve full space and a third GB at creation
– Over 1 GB is metadata; allocates in 1 GB slices as needed – Smallest allocation is 1 GB, contiguous address – Best for higher performance applications
• Note: metadata added as allocation increases
Many disks in pool
Thick LUN 100 GB
Thin LUN 100 GB
Reserve 99 GB Allocate 2 GB
Allocate 3 GB
99 GB RESERVED
3 GB
2 GB
19 © Copyright 2012 EMC Corporation. All rights reserved.
Pool Structure Private RAID Groups are created to host the virtual LUNs
– RAIDgroups will hold drives of one technology – SSD, SAS, NL-SAS – Different sizes and speeds can be combined, best if not
– Preferred disk count: RAID 5 multiples of 5, and RAID 6 and 1/0 multiples of 8
– Zero and Verify when pool is created (or when slice is freed)
LUNs carved from hidden RAID groups – Depends on number of drives – LUNs = 2 x (Pool Private RAID Group Disk Count)
The 1 GB Slice – Like a very thickly sliced metaLUN: data in that slice is serviced from a
single RAID group. Imagine a metaLUN with a stripe multiplier of 4096
Tracking Granularity – Thin LUNs: 8 KB unit of allocation – Thick LUNs: 1GB unit of allocation and effective stripe element from each
underlying RAID group
20 © Copyright 2012 EMC Corporation. All rights reserved.
Pool Structure: RAID 5 example
10 disks 2 Raid Groups (4+1), private (not shown in GUI)
From each RG: 10 LUNs – private; Half to SPA, half to SPB
Virtual LUN hosted on SPA
SPA
SPB LUN 12
SAS
SAS
SAS
SAS
SAS
RAIDgroup 1 RAIDgroup 2 SAS
SAS
SAS
SAS
SAS
Slices are taken evenly from all RAIDgroups in the pool
21 © Copyright 2012 EMC Corporation. All rights reserved.
RG-1 FLU-10 RG-2 FLU-20
1st 1GB slice
2nd 1GB slice
3rd 1GB slice
4th 1GB slice
nth 1GB slice
(n-1)th 1GB slice
1st 1GB
2nd 1GB
4th 1GB
6th 1GB
3rd 1GB
5th 1GB
50th GB
User accessed space: undetermined. Could be as little as 24 KB! User allocated space, 3 GB. Physical space consumed Total 6 GB
LOG
ICAL
PHYS
ICAL
LBA-0 LBA-2097152
Example, pool with single 50 GB LUN
If any part of this 1GB LBA ordered slice is written to, the whole 1GB slice will be allocated
Virtual space – reserved but not allocated Physical pool space – unallocated yet
LUN 1 User Data
Metadata
Legend
RG-2 FLU-21 RG-1 FLU-11
Pool Structure: Allocation
22 © Copyright 2012 EMC Corporation. All rights reserved.
Pool Structure: Allocation
1st 1GB
2nd 1GB
4th 1GB
6th 1GB
3rd 1GB
5th 1GB
last 1GB
Example: Pool with multiple LUNs
LUN 1 LUN 2
Virtual space – reserved but not allocated
Physical pool space – unallocated yet
LUN 1 User Data
Metadata
Legend
LUN 2 User Data
LOG
ICAL
PHYS
ICAL
RG-1 FLU-10 RG-2 FLU-20
1st 1GB slice
2nd 1GB slice
3rd 1GB slice
4th 1GB slice
nth 1GB slice
(n-1)th 1GB slice
RG-2 FLU-21 RG-1 FLU-11
LUN 3
LUN 3 User Data
Note that data is interleaved, but concentrated at the ‘bottom’ of the RAIDgroups – unlike old-style FLUs which would have the disks seeking large distances between each LUN start LBA.
23 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP Mechanics: Policies
• Three policies control movement of slices (+ No Movement) • These can be used to finesse your storage if needed • But first we should learn a bit about how they work…
Tier Policy effect at slice allocation Tiering Policy effect at relocation
Auto Tier Distributed across tiers randomly
Auto Tier Moves data according to temperature
Highest Tier Available Place as much data as possible on
highest tier
Highest / Lowest Tier Available Moves data into preferred tier as
possible; can move down
Lowest Tier Available Place as much data as possible on
lowest tier
No Movement No data movement after initial allocation New allocations will follow tier policy set
at LUN creation
24 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP Mechanics
• ‘Auto’ allocation is for general purpose, ease-of-use
– Slices allocated in random distribution across all tiers
– Distribution of new slices is in proportion to capacity of each tier
SPA SPB
CORE CORE CPU CACHE CORE CORE CPU
CACHE
SAS SAS SAS SAS SAS SAS
25 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP Mechanics
SPA SPB
CORE CORE CPU CACHE CORE CORE CPU
CACHE
SAS SAS SAS SAS SAS SAS
• ‘Auto’ allocation is for general purpose, ease-of-use
– Slices allocated in random distribution across all tiers, weighted by free capacity
– Distribution of new slices is in proportion to capacity of each tier
• Tier relocation is initiated by temperature and space
– Relocation of Auto LUN slices cannot allocate all of an upper tier (some held in reserve)
26 © Copyright 2012 EMC Corporation. All rights reserved.
FAST VP Mechanics
• ‘Auto’ allocation is for general purpose, ease-of-use
– Slices allocated in random distribution across all tiers, weighted by free capacity
– Distribution of new slices is in proportion to capacity of each tier
• Tier relocation is initiated by temperature and space
– Relocation of Auto LUN slices cannot allocate all of an upper tier (some held in reserve)
• Improve determinism with “Highest Tier” policy
– ‘Highest’ slices get priority over any ‘Auto’ slice
SPA SPB
CORE CORE CPU CACHE CORE CORE CPU
CACHE
SAS SAS SAS SAS SAS SAS
‘Highest’ priority ‘Auto’ slice will not move up if top close to full
Example: highest tier near full
27 © Copyright 2012 EMC Corporation. All rights reserved.
Tiering Summary Watch your CPU (keep below 60%) when migrating to FAST Tiers
FAST will use resources, and if EFD are in the tier, allow higher rates – but that will require higher CPU usage
Adjust watermarks down to make up for DRAM cache lost when FAST is installed
AUTO TIER is best for – Moderate-performance LUNs -- because new slices are busiest, and mission-critical apps should
have more deterministic allocation – TCO: use with SAS and NL-SAS – Simplicity
High-performance environments will take a little more planning – Use Highest, Lowest, No Movement options for data that clearly requires prioritization
Manage your migrations into a pool
You may need more than one pool – Multiple failure domains: don’t put logs & data in same pool, etc. – Pools with FAST Cache and those without (Fcache is a pool-based attribute) – Separate unlike workloads (sequential from random etc.)
28 © Copyright 2012 EMC Corporation. All rights reserved.
Conclusion
Know your workload
Size the array according the workload
Balance across every available component
Use Performance Health Check to confirm that your array setup is correct