re-think storage – pernixdata. meet & greet with frank denneman
TRANSCRIPT
© PernixData. All rights reserved.© PernixData. All rights reserved.
Re-Thinking Storage by Virtualizing
Flash and RAM
Frank Denneman, Chief Technologist
PernixData
Traditional I/O path in virtual datacenters
© PernixData. All rights reserved. 2
Scale out-compute using single storage array
Context based I/O inside host, loss of application identity outside host
Non integrated QOS throughout stack
Oversubscribed data-path
Solution: add more disks to array
Switch Switch
VM VM
ESXi Host
Storage Array
VM VM
ESXi Host
Flash placement
© PernixData. All rights reserved. 3
Server Array
Network
But flash performance is all about distance
© PernixData. All rights reserved. 4
Stage 1 Stage 2 Now
Industry development
© PernixData. All rights reserved. 5
Flash behind HBA Flash on PCIe
Flash on DIMMs
Flash-RAM convergence
RAM
Lightspeed Ludicrous speed
Million IOPS for 5K USD
© PernixData. All rights reserved. 6
Leveraging flash performance in your architecture
© PernixData. All rights reserved. 7
Insert
Flash
HereDiagram by Chad Sakac
© PernixData. All rights reserved.
ARCHITECTURE
OVERVIEW
PERNIXDATA FVP
PernixData FVP
© PernixData. All rights reserved. 9
Accelerating
application
read and write I/O
You are in control!
© PernixData. All rights reserved.
PernixData FVP clustered platform
© PernixData. All rights reserved.
Supports mobility at scale
ESXi Host
Flash
Device
ESXi Host
Flash
Device
VM
Remote flash accessFVP
(DRS)
Migration VM
Read
Storage Array
data
Write through policy
© PernixData. All rights reserved. 12
Write to Flash & Array
I/O complete after ack
of both resources
Primarily accelerating
subsequent reads
Offloads array & SAN of
Read I/O
Write back policy
© PernixData. All rights reserved. 13
Write to Flash then ack
Write to Array
transparent to
application
Read & Write
acceleration
© PernixData. All rights reserved.
Fault tolerance
© PernixData. All rights reserved.
FVP Fault Domains
Fault Domain
Blade Center 2
Fault Domain
Blade Center 3
Fault Domain
Blade Center 1
vSphere
Cluster
Storage Array
VMVM VM VMVM VM
Acceleration
resources
ESXi Host
5
Acceleration
resources
ESXi Host
6
AVM VM VMVM VM
Acceleration
resources
ESXi Host
3
Acceleration
resources
ESXi Host
4
VMVM VM VMVM VM
Acceleration
resources
ESXi Host
1
Acceleration
resources
ESXi Host
2
Peer host selection Peer host selection
FVP
© PernixData. All rights reserved.
INFRASTRUCTURE
ALLEVIATION
© PernixData. All rights reserved.
Accelerate VM = The most social thing you can do
8 Billion Reads didn’t reach primary storage
318 Terabytes of storage bandwidth not used
10 Billion Writes saw significantly low latency
Freeing up resources for other workloads
Hero numbers in 6 weeks time:
© PernixData. All rights reserved.
Datastore
writes
Application
performance
Uniform write patterns
Smooth out writes
Time release I/O
Reduce forced array cache flushes
Scale out
© PernixData. All rights reserved. 19
16 VMs running SQLIO
Work Profile: Series of sequential reads and writes at 8K, 16K, and
64K, then random reads and writes of same block size pattern
$2.5 million USD array (64 EFD’s, multiple TB of cache)
FVP+SAN
1.6 Million
IOPS
Array:
51000 IOPS
© PernixData. All rights reserved.
Offload your storage controller CPU
Setting up and maintaining data paths
Mirror writes (storage controller redundancy)
Data movement and data integrity
Maintaining RAID levels, calculating & writing parity data
Snapshots, Replication, Dedup, Compression
Multi-tiering algorithms
Running integrated management software
PernixData FVP: Recovering useable capacity
© PernixData. All rights reserved. 21
CPU at Max (~95%), over 200TB of stranded storage
FVP offload array utilization (IOPS and CPU) by 75%
Releasing Stranded capacity
© PernixData. All rights reserved.
TODAY’S DATACENTER
WORKLOADS
DEMANDS A
NEW ARCHITECTURE
© PernixData. All rights reserved.
Plan-able performance scale-out
Non disruptive performance scale out
Linked Compute + Storage performance growth
Future proof: new tech instantly usable
Decouple storage perfromance from capacity
© PernixData. All rights reserved. 24
VMVM VM VMVM VMVM
FVP
ESXi Host
Flash Device
ESXi Host
Flash Device
ESXi Host
Flash Device
Storage Array
VM VM
Performance
Capacity
+
Data services
Deterministic performance
© PernixData. All rights reserved. 25
Today’s possibilities
© PernixData. All rights reserved. 26
New gen. server media
…
FVP
+
32-192 TB RW acceleration
resource in 32 node cluster
250.000 IOPS per host
32 Million IOPS per cluster
Consistent sub millisecond latency
40Gb interconnects between hosts
An insurance policy for current and future workloads
© PernixData. All rights reserved. 27
© PernixData. All rights reserved.
Deterministic performance
Application performance
– Change perception of storage performance
– Consistent and predictable latency
Infrastructure alleviation
– Insurance policy for existing application performance levels
– Absorb any unknown workload without affecting current
workloads
Service offering
– Control which application gets performance
© PernixData. All rights reserved.
Questions?@frankdenneman
@pernixdata