1. Loris Giovannini, Mauro Giacchini Epics Collaboration Meeting
Archiver Hypertable Engined
(HyperArchiver)
Loris Giovannini, Mauro Giacchini
Epics Collaboration Meeting
03/06, 2010
L.Giovannini -- M.Giacchini

2. Target
Make the EPICS Archiver embedded DB more reliable and safe especially in large application with a large number of PVs to store
First Approach: re-write the Rtree and double- linked-list embedded DB structure
Second approach: look for a complete replacement of the embedded DB with new one
L.Giovannini -- M.Giacchini

3. Why not Beauty ? (Best Ever Archiver Utility)
Beauty = Classical Archiver over Oracle DB
Pros:
SQL statements accept
Professional Solution
Cons:
Not so fast
Really expensive (licenses)
L.Giovannini -- M.Giacchini

4. Why not MySQL ? Potential scalability concerns
Designed for a single machine (not distributed)
What it takes to make it scale
Major engineering effort
Solutions are usually ad hoc
Solutions usually involve horizontal partitioning + replication
Solutions involve expensive hardware
L.Giovannini -- M.Giacchini

5. BigTable ?
Proprietary DB used by Google over their proprietary Google File System
High Performance
Large amount of data
Stable
Reliable
Distributed
L.Giovannini -- M.Giacchini

6. HYPERTABLE…what is it ?
Hypertable has been developed as an in-house software at Zvents Inc. In January 2009, Baidu, the leading Chinese language search engine, became a project sponsor. Rediff.com is one of the premier worldwide online providers of news, information, communication, entertainment and shopping service.
Hypertable is licensed under the GNU General Public License Version 2.
“Our goal is nothing less than that Hypertable become one of the world’s most massively parallel high performance database platforms.”
Not RDB
L.Giovannini -- M.Giacchini

7. Hadoop/Hypertable Solution Objective
Handles applications with large datasets
Fast fault detection and fail-over
High throughput processing
Centralized scheduling of tasks and execution of batch processes
Can be deployed on low cost, commodity hardware
Eliminates or reduces the need for expensive table joins
L.Giovannini -- M.Giacchini

8. Machines used to the tests
Machines and time up
Insertion Time
Retrieval Time
(all pv parameters)
Virtualbox virtual machine Debian Linux Stable x86 4 cores (shared among 3 VMs) 4GB RAM (1) 5.4k RPM SATA Disk $4k (to 4 Vbox)
(1/2 h data acq SNS)
1 min 31 sec
sample
19.5K samples/sec
0.042 sec to extract
133K samples of 4 pv
3.17M samples/sec
Debian Linux Stable AMD cores 64GB RAM (7) 10k RPM SAS Disks (RAID 5) $11k
1min 13.6 sec
samples
24K samples/sec
0.024 sec to extract
133K samples of 4 pv
5.54 Msamples/sec
Debian Linux Stable AMD cores
64GB RAM (7) 10k RPM SAS Disks (RAID 5) $11k (7dy data acq SNS)
11.2 hours to insert
597M samples
14.8K samples/sec
0.031 sec to extract
samples of 4 pv
4.3M samples/sec
19min 51 sec
all pv
500K samples/sec
64GB RAM (7) 10k RPM SAS Disks (RAID 5) $11k (30dy data acq SNS)
37.86 hours
2557M samples
18.7K samples/sec
0.030 sec to extract
L.Giovannini -- M.Giacchini

9. Oracle Vs Hypertable
Oracle insertion time : peak in write test (8Ksamples/sec)
Hypertable insertion time: per 1min 13 sec (24Ksamples/sec)
Oracle retrieve time: samples in sec (~462 vals/sec)
Hypertable retrieve time: samples in 0.36 sec ( 160K vals/sec) in half hour
L.Giovannini -- M.Giacchini

10. Insertion Time Hypertable
Hypertable into CSS
THE OUR GOAL
storing: channels per second
retrieve: 1000 samples for up to 4 signal in less
than 1 second
Machines and time up
Insertion Time Hypertable
Insertion Time
MySql
Ubuntu 9.10 karmic Kola Intel Pentium M 740 1,73Ghz 2GB RAM 5,6K RPM sata Disk
8,45 sec
sample
13K sample/sec
27,1 sec
110200
4K sample/sec
L.Giovannini -- M.Giacchini

11. Retrieval outside CSS First Step:
Develop a Java Program for only test the performances for retrieval without any
elaboration of data to valuate the possibility to respect our goals using Thrift Java client API (Mandatory to interface with Hypertable)
Results 1000 sample for 4 Pv:
1000 sample:22ms
1000 sample:15ms
1000 sample:15ms Ksample/sec
1000 sample:19ms
L.Giovannini -- M.Giacchini

12. Retrieval
L.Giovannini -- M.Giacchini
L.Giovannini -- M.Giacchini

13. TIME Retrieval into CSS
Second Step:
Develope an Eclipse plug-in org.csstudio.archivereader.htrdb for the support of
Hypertable like as org.csstudio.archivereader.rdb
Working in Progress
The extraction time into CSS is about 10 times worse:
The reasons are under study, there are two probable causes:
initial query latency of Thrift API when connect to
Hypertable
wrong classes design into CSS architecture
L.Giovannini -- M.Giacchini

14. Acknowledgments More information Kay Kasemir (ORNL)
Doug Judd (Hypertable Core Team)
Bob Dalesio, Ralph Lange, Robert Petkus (BNL)
More information
L.Giovannini -- M.Giacchini