2. A bit about myself
Eyal Markovich
Senior Director of Performance
• 20 years APM and database performance experience
• 18 years in Precise working with hundreds of
customers
• 4 years in Kaminario managing the performance
team
3. Performance Analyses Learnings:
1. Business-critical applications require flash
2. Not all flash products are created equal
3. They Require 3D performance consistency
4. More than 80% of the
performance issues are traced
back to poorly performing
storage
5. Let’s start with an example
• Large financial institute
• Slow Batch job run for 7 hours
• Often exceeding batch window
• Critical data not available for the business
6. 20 years: similar disk technology
Selecting 1% of the database …
Performance:
5X
1990 1994 1998 2003 2011
Data:
1400X
1
Min
2
Min
3.5
Min
18
Min
11.5
Hours
1.4 GB
4 GB
9 GB
73 GB
2 TB
7. Storage Architecture Evolution
XIV
- No shared
storage
- Over provisions
$$$
- No shared
workloads
- Difficult to grow
- Can’t scale easily
- Under utilized
capacity
- Creates silos
- Hard to manage
-Disrupted
Performance
Limited
-Doesn’t fit flash
-Performance
Limited
- Same as DAS
- Tuned for single
workloads
- Real Time,
elastic and agile
- 3D performance
- IOPS focused only
- Resilient
- Inconsistent
- Scalable
- Not resilient
- Best TCO
- Not scalable
8. Flash – Competitive Landscape
3D Performance
Silo’d ($$$ & Resources)
Limited Workloads
Limited SW stack or 3rd party
Limited Scalability
No enterprise HA
Silo’d ($$$ & Resources)
One Workload
No SW stack
No Scalability
No enterprise HA
Mixed Workloads
Full SW stack
Easiest management
Best TCO
Enterprise HA
Silo’d ($$$ & Resources)
Limited Workloads
Limited Scalability
SLA under failure
Software
Capabilities
9. Back to the example
• Large financial institute
• Slow Batch job run for 7 hours
• Often exceed batch window
• Critical data not available for the business
I/O Wait before
Kaminario
I/O Wait
After Kaminario
CPU
10. What about my Oracle instance?
• Do I need Flash for my Oracle?
CPU
Lock Net
I/O Waits
8:00 am
• What type of Flash is best for me?
• SSD cards? Hybrid array?
• SAN flash array? Scale out SSD flash array?
9:00 am
11. What about my Oracle instance?
• Do I need Flash for my Oracle?
12. What will be the effect of Flash on my workload?
13. What will be the effect of Flash on my workload?
db file sequential read
14. What will be the effect of Flash on my workload?
db file scattered read
15. What will be the effect of Flash on my workload?
Direct Path Read
Sample SQL Query:
Select /*+ Parallel(emp 4)
*/ * from Employee emp;
Execution Plan:
SELECT STATEMENT
PX COORDINATOR
PX RECEIVE
PX SEND RANGE
PX BLOCK ITERATOR
TABLE ACCESS FULL
EMPLOYEE
16. What will be the effect of Flash on my workload?
Log file sync wait
17. What will be the effect of Flash on my workload?
Direct path read temp and direct path write temp
18. Performance Analysis Issues:
•
•
•
•
•
•
•
Random read latency for small block size
Large latency for sequential reads
Mixed workload applications running both OLTP & reports
Parallelism set too high or too low
Writes (DML) affecting read latencies
Storage unable to support temporary workload
Large latency for writes
Inconsistent Performance Across Workloads
19. SAP Example: Simple Workload
Two Distinct Workloads:
1.
2.
Dialog User activity
Programs (batch)
Running Overnight
• Large sequential scans
• High parallelism on the database
• High throughput
20. SAP Example: Simple Workload
What we need from storage to improve the
performance:
LOW
Throughput
1,483 MB/s
HIGH
IOPS
4,410
Latency
0.89 ms
21. SAP Example: Complex Workload
Can you run the SAP programs during the day?
Activity:
• What will the effect be on User activity?
• How well will the SAP programs perform?
It’s All About Storage:
• What will the effect be on storage latencies?
• What will the effect be on storage throughput?
• Will storage performance remain consistent ?
Can your system tolerate mixed workload?
22. SAP Example: Complex Workload
You need flash that supports mixed workloads:
•
•
Consistently low latencies regardless of the throughput
Delivering high throughput regardless of the IOPS
LOW
Throughput
1,713 MB/s
HIGH
IOPS
24,426
Latency
.56 ms
23. What are Kaminario K2 differentiators?
Storage done better
Scale-Out
3D Performance
Granular & Linear (6120 TB)
Consistent
Automatic
performance &
endurance
Linear scale (200K- 2M+
IOPS, 3-30 GB/s)
7 year flash endurance
Simple & Intuitive
Global management
Low TCO than
alternatives
Lowest Latency (280 us)
Mixed workload
concurrently
Minimal performance
degradation (<25%)
during recovery
Safe & Reliable
No Single Point of
Failure
Self healing – N+1
Instantaneous, space
efficient, snapshots
Non disruptive
upgrades
Full node recovery
~ 90 minutes
24. Kaminario K2 Building Block
Enterprise Grade Highly Available Building Block
Cluster together to scale-out Performance or Capacity
Feature
Flash
K-Block
K-Node
Specification
8 MLC Flash (800 GB or 1.6 TB) per K-Node
Usable Capacity
3 TB / 6 TB per K-Node
(6 – 12 TB) or (12- 24 TB) per K-Block
Performance
IOPS - 200K-410K
Throughput - 3–6GB/s
Latency - 280 us
Footprint
6 RU – 8 RU
Power
Typical 1.4kW-2kW
High Availability
System components fully redundant
Dual Processors, dual battery
Host Ports
2 host ports per K-Node
8 x 8Gbp FC HBA or 8 x 10GbE iSCSI per K-Block
K-Node
K-Node
K-Node
29. SPC-1 2013 Configuration
CONFIGURATION
•
•
•
•
86 TB System
7 K-Blocks
34U Footprint
SMART – All Flash
CONFIGURATION RESULTS
• 3 Million Read IOPS at under 2ms latency
•
2.8 Million Read IOPS at under 1ms
•
1.2 Million Write IOPS at under 1ms
• 56 GB/s Read/Write Throughput
30. 3D Performance: Latency under failure
Kaminario Scale-Out Flash Array
Hybrid & Monolithic Flash
Additional info we’ve found when doing these performance assessments
SPEAR minimizes writes to the Flash media and maximizes Flash endurance.It utilizes a system-wide write cache that eliminates hotspots, and it distributes writes across all Flash media in the system, thus sustaining global wear leveling. SPEAR’s efficient management of metadata eliminates writes to the Flash during updates and facilitates space efficient snapshots that do not incur additional writes. K2 deploys enterprise-grade level SSDs with advanced Flash management capabilities that optimize the endurance of theFlash . With SPEAR’s abilities, Kaminario has a warranty of 7 years for Flash enduranceWe use enterprise SSDs with advance error correction techniques and with optimization related to the way the data is written to the Flash drives.
The K-Node is build from 4 1U K-Servers. Super Micro ServersThis unit is best optimized for performance, and the performance is in the ball park of these figures.Now, you just need so scale that performance and capacitySpeak verbally performance enhancements % improvement vs. V3 . Even though cost effective Enterprise SAS MLC – IOPS and Throughput much better. Latency bit lower but still within <0.5 ms target. Also talk about performance during recovery – how using backup to Flash meets customer need for performance even during worst times!The K-Node is build from 4 1U IO serversThis unit isbest optimized for performance, and these are the figures.Now, you just need so scale that performance and capacityLet’s talk about V4’s HW architecture.We use a 1U rack mount architecture, which allows us to benefit from good footprint and power, as well as high media density.The servers are based on the latest and best processing power available today, and our advantage here is that we can easily benefit from any progress made in this area.Since we are not bound to a single vendor, we can always benefit.As said before, the K2 is an all SSD array, where the primary and secondary media are MLC based SAS Solid State Drives.Our system is vendor agnostic, and we are implementing a dual vendor strategy. One of our vendors is SMART. (SECOND VENDOR: if asked, we are still in the works of deciding on a second vendor)(NAND: if asked, the NAND supplier of the drives we are utilizing is Toshiba)The spare node, of our N+1 design, is consolidated together with the management servers, which improves foot print, power and cost.The spare is used only in the scenario of a recovery, and IOs are always served from an SSD device.We are currently deploying 8 drives of 800GByte per 1U, with the intention of deploying 1.6TByte drives per 1U.Also here, we can benefit from market changes.(SHAI: the MLC is a 2X technology, 24nm, 10 writes per day per block, full capacity for 5 years) Optimus™ SAS SSD There is no UPS in the system, rather each server has a battery that insures a graceful shutdown upon a power failure, securing the data.Our network backbone is a fully redundant 10GigaBitEtherent and we support FibreChannel and iSCSI protocols.The K2 is managed from the centralized K-MGMT block, no matter how you scale.It contains two redundant MGMT Servers, and two redundant 10GigaBitEthernet switches.It is in charge of managing all aspects of K2: create volumes, monitoring, external connections and more.EXPLAIN: Network usage for internal traffic onlyDataControl (MGMT)
Take the lower red to just below 2 and the upper line to 20 ishAdd performance logo guaranteed5X and goes to 10-15x