1. Intel® SSD DC S3700 Series
Consistency Matters
Rob Crooke
Corporate Vice President, General Manager
Intel Non-Volatile Memory Solutions Group
2. Data Center: Environment is evolving
Path to Exascale:
Efficient for
Highly Parallel Applications
+
Parallelism does the work to extract real
Insights from Big Data
3. Data Center: Environment is evolving
Cores must be supplied with data continuously
to achieve optimal performance and results
+
9. High performance data centers require a new
generation SSD
Slow
Fast
Very
Fast
Consistency matters as much as IO performance
10. Intel® SSD Solutions for target markets
Intel® SSD Data Center Family
Intel® SSD Professional Family
Intel® SSD Consumer Family
11. Introducing Intel® Solid-State Drive DC S3700 Series
Consistently Amazing
Fast and Consistent
Performance
High Endurance
Technology
Stress-Free Protection
12. Intel® SSD DC S3700 Series features
the 3rd Generation Intel-developed controller
Utilizes continuous background cleaning
13. BIG DATA
Intel® SSDs accelerate a broad range
of Enterprise Applications
HPC
CLOUD
COMPUTING
14. The Intel® SSD DC S3700 Series Story comes to
life:
Every micro second counts
Introducing Mike L. Norman, Ph.D, SDSC Director
15. Whole Database Protein Pairwise Alignment Calculation
Speedup of 4x when using Flash-based I/O Node
Intel SSDs Provide 4x Speedup over
Traditional Disk Architectures
16. Prediction of New Catalysts for Olefin Metathesis
Intel® SSDs support a class of HPC that has
not been possible previously
Z-selective olefin metathesis transition state
17. Accelerating historical data analysis in capital
markets
Introducing Peter Lankford, STAC® Founder / Director
TM
18. Intel® SSD DC S3700 Series:
Consistency Matters to utilize the heartbeat of computations
1-2 minutes
Highlight two overall themes of the talk: Consistent performance in Data Centers & introduction of the S3700 Series SSD
Rob to introduce himself
We’re going to tell you why Consistent performance in the Data Center matters
Introduce an exciting new Intel Data Center SSD that delivers consistently amazing results
2-3 minutes
As Diane mentioned before me, the Xeon Phi launch takes Data Center processing to the next level. It allows for HPC applications to do more with more cores than ever before.
Compute the fastest route in traffic based on mobile phone gps data of people in cars.
Predicting Behavior of Urban Infrastructure during Disaster
Compute and model genetic mutations in Cancers from Genomes of patients worldwide.
Predicting Response of Customers to Product Design
HPC Creates High Value Insights From Big Data (HPC is the Mint.. Big Data is the paper) (HPC is the Pump, Big Data is the Old Field, Insight is the Oil)
3-5 minutes
The cores need a continuous feed of data to be fully utilized
Xeon Phi tie in. Introduction of many threads
Feed the cores to utilize processing power (FOIL: Storage feeding the cores, maybe same Foil as 2a)
Therefore, Need for lower latency increase bandwidth storage to feed cores/threads
So the right solution is an SSD? Partially correct.
5-6 minutes
Now that we have established the need for data with our new Data Center processors, then this need can be satisfied with SSD solutions on the market today. In previous launches, Intel has highlighted that not all SSDs are created equal. This is certainly the case in architecture, however in recent years, the performance across SSDs have become increasingly similar.
6-7 minutes
SSD Environment
SSDs don’t have consistent performance. Based on their workloads and wear leveling, they can have inconsistent results per iops over time
Most SSDs Stalls computation
Typical SSD has more variability than a HDD
7-9 minutes
To bring this back to consistent performance, This graph shows the performance data over the entire 2000 second test period. In this charts you'll notice an early period of very high performance followed by a sharp dropoff. What you're seeing in that case is the drive alllocating new blocks from its spare area, then eventually using up all free blocks and having to perform a read-modify-write for all subsequent writes (write amplification goes up, performance goes down).
Everytime the IOPs fall, the processor cores are under utilized.
Data Center administrators may design their solutions around the worst case IOPs (Raid Control, & Striping Data)
9-10 minutes
Hard Drives. You can site that they are more consistent, however the IOPs re no where close to any sizeable performance level
HDD are more consistent but not as good with performance
When you’re that slow, the entire range make look/seem consistent
10-12 minutes
Hard Drives are not as volitale, however will still be the bottleneck for data distribution
12-14 minutes
The opportunity exists to provide Very Fast IOPs consistently over time. Intel recognized this opportunity and have been working the last three years to build an SSD from the ground up to address consistency, along with the performance and reliability central to all Intel solid-state drives.
Consistency matters as much as IO Performance to meet the needs of the cores, utilizing Xeon Phi’s capabiities.
14-15 minutes
Intel SSDs target three key target markets
15-17 minutes
Hold up a sample unit, allow time for pics.
Verbally call out specs and interweave terms: Size, Complexity, Accessibility.
“Intel believes that a 4 sigma availability represents the max latency requirements needed in most datacenter applications”
Fast and Consistent Performance
Deliver data at a breakneck pace, with consistently low latencies and tight IOPS distribution.
75K Random Read IOPS1
Latencies: Typical 50µs; Max <500µs2
Stress Free Protection
Protect your data center applications with multiple secure checkpoints that provide protection against data loss and corruption.
Full data path and non-data path protection
Power safe write cache with built in self-test
High Endurance Technology
Write endurance of 10 drive writes per day over five years
Complete feature set (Pete comment)
23-24 minutes
Dirty Dishes. We all have to deal with dirty dishes. Every time a dish is dirtied, you make a decision on whether to clean it immediately or let it wait until a more convenient time, possibly allowing dishes to pile up in a sink. Most SSDs wait until the “sink” is full to clean blocks. When this full sink cleaning occurs, latencies can increase significantly thus stalling system performance.
The new controller on the S3700 cleans blocks early and often, not allowing data to queue up and keeping the flow of data consistent to the processor.
Data movements for read disturb or wear leveling. Read disturb, over time if you read a cell next to another cell it can cause the cell to be overprogramed, which may result in data loss. Proactively move data after a certain # of reads or when data is read, you know how many errors that you have, if it’s close to that threshold, we move the data.
Data is moved with power safe. Move first and then erase data.
WHAT MAKES IT DIFFERENT: How its different, the thresholds and the data we pick and how often we do it.
Incorporates the learning from the previous generations, and feedback from customers like
4 out of every 10 SSDs going into data centers …
Tech rep: Continuous background defrag, as needed. Meaning, after everything is clean then the defrag stop. It starts when a certain amount of stale data accumulates, based on firmware thresholds. Stale data is duplicate that is old and invalid.
24-27 minutes
Big Data –
The S3700 enables faster processing of complex queries / analytics on concurrent jobs;
Fewer nodes are required when using the S3700 compared to a cluster using HDDs which helps drive down the total cost of ownership of the system.
Using the S3700 meets the high IOPS needed during mapping and sorting operations
Cloud Computing –
Datacenters are struggling to provide the high level of service there customer expect with their existing solutions. Traditional hard disk drives (HDDs), spinning disks with mechanical parts, are stretched to their limits by virtual machines (VMs) that demand high levels of random input/output (I/O) performance. The S3700 , with no moving components, is better equipped with its fast and consistent to perform the intense I/O required by multiple Virtual Applications.
SSDs bring significant value to Cloud Computing
A sample configuration, tested by Intel’s IT Research organization, demonstrates a multi-VM random workload achieving 90-100K I/Os per second (IOPS) using only 24 Intel® Solid-State Drives, a typical cloud workload that would require 450 traditional 15k RPM HDDs to accomplish the same results.
Results of the test demonstrated that a US $43,000 solution is capable of supporting a continuous stream of 90-100K IOPS with a 100 percent random workload at only 550 watts of power and 2U of rack space. OEM solutions with HDDs would typically cost $200,000 (based on pricing at the time of the white paper in early 2012).
High Performance computing – from SDSC paper
“With outstanding performance, Intel® SSDs help SDSC’s researchers perform ground-breaking scientific research. Flash provides an order of magnitude improvement in I/O over traditional HPC disk architectures.”
Michael Norman
Executive Director
San Diego
Supercomputer Center
Gordon uses Intel® SSDs to provide a new level in the memory hierarchy. This is enabling new classes of applications and scales of computation not previously possible on HPC architectures.
22-29 minutes
High level introduction
Here’s just one example of what I’m talking about. This is from the Protein Data Bank (PDB), which is the single worldwide repository of information about the 3D structures of large biological molecules. These are the molecules that are found in all living organisms. The 3-dimensional arrangement of each of these proteins provides the complexity and richness of life as we know it. Comparing proteins to one another is critical to many life science domains, including drug discovery. The process of comparing the tens of thousand of proteins in the PDB can literally take months to complete. By moving the database to SSD, we have seen more than 4x speedup when compared to traditional disk-based architectures. So a complete pairwise comparison can now complete in weeks instead of months. This kind of performance improvement accelerates scientific discovery and is not possible by simply adding more flops to a traditional HPC cluster.
Recently, we tested the new Intel S3700 drives in on of Gordon’s I/O nodes and were absolutely thrilled to see an additional 2x improvement in speedup over the current Intel 710 drives that are deployed in Gordon.
Supporting a class of HPC that has not been possible previously. More and more user demand. Need that consistent performance.
Energy calculations for bonding
Quote from Kendal: Using computable method is the gold standard.
Useful in Consumer product generation
Use quote in slide
Use flash for extra crash.
Abacus
Flash in the node. No local storage for HPC.
Move closer to the processor
Olefin metathesis is a powerful tool to form carbon-carbon double bonds in organic synthesis and materials
chemistry. In collaboration with the Grubbs laboratory at Caltech, we are using Gaussian calculations to predict the selectivity of new ruthenium olefin metathesis catalysts. Using flash disk and large amount of shared-memory has dramatically increased the performance of the calculations and enabled more accurate computational methods such as CCSD(T).
Computations of hundreds of transition states with ~100 atoms using high-accuracy DFT and CCSD(T) methods.
Frequent I/O operations on 100GB+ scratch files
Large memory requirements (2GB/core)
Description of the STAC. Importance / impact of STAC
29-31 minutes
We’ve already discussed the goodness of the flow of data between the S3700 Series and Intel’s Processors. To make this correlation more clear, I’d like to call up Dr. Michael Norman to the stage.
