There is a need to accelerate the performance of database deployments that are used for general e-commerce or enterprise application purposes, in both scenarios where transactions are high and where database is used for information retrieval. It is very tedious to fine tune the database performance.
Database today does not have storage management policies that are aligned with the needs of the database workload pattern. Once the information is not available inside the system memory, the database has to make the costly round trip from the back end storage where it is further slowed down because of disk access. The network becomes as slow as the disk on the back end. On the storage back end there are no policies available to fine tune the database.
Now Datagres has found a technique to align the storage for the needs of the databases. The good news is that now Database and Storage administrators can align the data management to achieve another level of performance in existing deployments.
Fact: PerfAccel software provides acceleration for Database environments from 3X-15X
While most database data is stored on the SAN, NAS, iSCSI, or DAS appliance – PerfAccel helps servers keep data close to the application. The key to PerfAccel’s technology is intelligence. Through detailed application IO analytics and policy based IOPS provisioning, PerfAccel empowers users with a never before seen level of “active data” management. By identifying trouble areas, users are able to fine tune IOPS policies over directories, tables, index files, and various databases. PerfAccel is an essential tool for any administrator running Hadoop, Oracle, MySQL, and PostgreSQL.
|Sequence Scan||Upto 27X|
|Index Scans||Upto 12X|
|Hash Joins||Upto 10X|
|Nested Loop Join||Upto 10X|
Database server relies mostly on the performance of the back end storage for reliability, performance and scaling perspective. The network storage has disks inside it and that causes the performance even on fast fiber channel network to be only a fraction of the bandwidth.
Every round trip causes delays and cpu has to wait for the data to come in.
Scaling the network storage is very expensive and only marginally improves the performance.
Datagres PerfAccel has dynamic data management control over the data in the cache and in the network storage.
The hot data can be cached close to the server. Administrator can define polices over what kinds of data stays close to the server.
With upto 70% reduction in the round trip to the back end storage. The database starts to perform 3X to 10X better.
Using Datagres Analytics one can figure out the active data size, the IO finger print of the workloads and decide on the polices that work best for your users.
Reduction of load on SAN, NAS
Upto 70% reduction of load on the back end Storage
Data Management Policies
Fine Tune Data Management Policies for your workload
Risk of Storage Downtime
Ability to Scale Database
Analytics to do Workload I0 Finger Print
Available for the first time
PerfAccel Deployment on the database servers just takes a few minutes. It needs the caching device on the system , that can be a local partition, or a whole disk or SSD or any other local device on the server.
Data source can be SAN, NAS or ISCSI or DAS.
The console can be deployed on a VM or another server.
No changes are expected in the database configuration.
There is a need to accelerate the performance of MySQL deployments that are used for general e-commerce or enterprise application purposes, in both scenarios where transactions are high and where MySQL is used to provide lookup services.
The typical popular usage of MySQL inside an e-commerce and enterprise application is close to the OLTP workload. In an OLTP type workload the database is used for recording transactions for an "incoming order" event. This scenario is fairly common across the internet where the transaction database is used to record the details of the online items being bought.
Another use of MySQL is for storing product catalogs or hosting wikis and knowledge repositories. These database instances are primarily read-only and tread the path of the data warehouse style of workloads.
As the volumes grow there is a big need to improve performance without buying new servers. This is because the performance is influenced by the IOPS speeds in the disks.
The OLTP workload is dominant with write requests and the data warehouse workload is dominant with read requests. However, it is not unusual to also find database instances that serve both types of workloads in equal measure.
Storage performance has been examined in the context of DW and it is challenging for DW to keep storage, memory and the disks working in a seamless fashion, since queries and data sizes keep changing.
Despite the standard workloads, it is important to understand that every single database instance has unique characteristics of its own and gets operated in very different business conditions throughout the day.
The good news is that Datagres PerfAccel provides hot data management policies for both OLTP and DW types of workloads.
SYSTEM UNDER TEST INFRASTRUCTURE DESCRIPTION
RESULTS TABULATION OLTP
RESULTS TABULATION TPC-H
The aim here is not to create another world beating OLTP, TPC-H benchmark but to study the IO pattern in the benchmark and observe if these patterns can be accelerated with Datagres PerfAccel Software in an intelligent way. The first generation of SSD caching software from other vendors does not distinguish between the IO of the OS, different parts of database and the logs.
Datagres PerfAccel Software provides unique abilities to apply different caching policies for different workloads on different data segments within a single database. Datagres PerfAccel also provides complete control over the directories, mount points and areas within the files, that can be selected for different acceleration policies.
ANALYSIS FOR THE OLTP
It is a usual problem that workload is not well understood. To understand the workload, the Datagres PerfAccel Software can be operated in the analytical mode only without the SSD. In this mode the Datagres PerfAccel Software will behave as if the cache is present without being actually present.
Following is the graph that summarizes the workload for the analytical workload.
From this graph, following observations can be made:
- There is a huge amount of read misses in the early part of the workload as data is not in the cache
- Once data is inside the cache the read hits start to go up. To circumvent this problem that is widely prevalent in most workloads, Datagres PerfAccel Software provides the pre-fetch functionality
- The workload is write centric and has enormous amount of write activity happening throughout the test, indicating that it could benefit from the write back cache
Based on these two simple observations, we ran the test with pre-fetch, loading the data as much as it can into the cache device and we also enabled the write-back. This time around the SSD is available in the system. Now the same graph appears as below:
Following observations can now be made:
Pre-fetching has reduced the time for the cache warm up to a very small negligible interval.
The number of read-hits per interval count has gone up enormously because of the cached data sets, and is scaling well as more read-hits can be achieved.
The write-hits from the cache have greatly accelerated the performance. The write-hits are also scaling well.
The entire workload finishes in a very short amount of time.
Further analysis of the workload is also available. The following graph shows the hot nodes and the amount of cached data from these nodes in the cache.
Datagres PerfAccel also gives you the unique ability to figure out the hot data points within each inode:
DATAGRES PERFACCEL SOLUTION ANALYSIS
CREST CASE STUDY
Crest’s render farm is comprised of 400 servers and 2 Network Attached Storage servers within a 1GbE network. Datagres PerfAccel analytics software module helped Crest discover NAS network IO was approximately 5TB per day.
OBJECTIVE: Deploy Datagres PerfAccel across 60 servers to offload IOPS from current NAS resources and reduce IOPS latency for render farm applications.
DEPLOYMENT: After the deployment of PerfAccel, Crest administrators discovered two-thirds of storage IO was serviced from the cache. The data read from the cache on a few jobs was near 90% of the data read requests of the system. Jobs completed 20% faster. These improvements saved over 1TB of network storage traffic within a three day period. More importantly, the NAS server performance was high and predictable.
PerfAccel software gave Crest administrators flexibility and control over their render grid IOPS without risk to data consistency and integrity. Through PerfAccel’s easy to use interface console, admins created cache policies for pre-fetching, predictive caching, real-time cache size configuration, and auto-caching of hundreds of NFS mount points. Another advantage of Crest’s deployment of PerfAccel was that users could use existing storage devices on servers as a cache device. Higher performance cache devices (SSD’s) could be added later as budgets allowed. PerfAccel provided Crest the ability deploy SSD cache devices without preference to vendor type or location.
SUMMARY: After 3 months, the investment of PerfAccel reached a break-even point thanks to these important factors:
With an estimated 5TB of high speed IOPS per day on tier 1 infrastructure, Crest spent approximately 0,000 per year for systems, maintenance, and management. Acceleration from local SATA drives provided a gain of 40 servers per year at an estimated cost of 0,000 annually.
The ability to employ a vendor agnostic infrastructure could only be measured over time with Crest’s growing need for additional storage capacity. Avoiding vendor lock-in has saved companies like Crest hundreds of thousands of dollars over the lifespan of their networks.
In conclusion, PerfAccel empowered Crest’s media render farm administrators with the ability to increase network productivity and overall application performance while lowering the costs of network storage with a visible return on investment within 3 months of deployment.
Proof of Concept (PoC)
At Datagres, we know that seeing is believing. Without disruption to your current infrastructure, Datagres will work with your team to deploy, analyze, and prove immediate return on investment using PerfAccel's technology.
PerfAccel – Intelligent management of Data in Motion
PerfAccel is a dynamic and intelligent grid management solution built tough for demanding industries like Media and Entertainment, ECAD, Big Data and enterprise level virtual environments. Through powerful analytics and reporting tools, PerfAccel creates smart, server side file level caching for storage IO and the management of thousands of nodes through a single pane of glass. With the ability to install invisibly on existing servers, Datagres PerfAccel deploys in minutes, without disrupting existing infrastructure. Once application IO requirements are analyzed and understood, PerfAccel dramatically reduces IO latency and offloads much (and often most) of network storage IO. Accelerate applications, scale, and manage entire grids with PerfAccel and get the most from your existing infrastructure.
Learn how some of the most data intensive users in the world rely on PerfAccel to accelerate data caching and improve network performance by up to 80%.
PerfAccel software delivers administrators flexibility and control to effectively manage data in motion while ensuring consistency and integrity. Through a simplified interface, storage administrators are provided with grid scale management software that enables analysis, reporting, and adjustment of IOP’s policies for predictive caching, real-time cache utilization, cache pre-fetching, and auto-caching of AutoFS mount points.
Once the performance bottleneck is alleviated from the back end NAS storage, network storage devices are measured in terms of the price of disk capacity not IOP’s. The need to continually upgrade expensive backend NAS storage for performance gains is reduced and in some cases, eliminated.