Abstract
This presentation was given at the 2000 International Symposium for Laboratory Automation and Robotics (ISLAR) held in Boston, MA, October 15-18, 2000. The full manuscript is available on CD-Rom and can be acquired by contacting Christine O'Neil, 508-497-2224; email
At Berlex Biosciences, laboratory automation has been utilized to assist in the pursuit of drug discovery goals. One system, a Zymark, has been developed to perform multiple compound formatting routines. The set up can add DMSO to vials containing research material, mix then transfer specified volumes of each solution to plate wells. Plates and vials are tracked using barcode readers to send the object ID to a storage variable. During the procedure, information is assimilated from a data input file specially prepared for the run, and two files are generated: a run log and a plate map data file.
The log is an ASCII text file. The contents are the title of the run, the start time stamp, any errors that occurred during operation, and the finish time stamp. The map file is a listing of the plate ID numbers with sample vial ID numbers and their corresponding positions on the plate. The layout is the SBS standard ninety-six well plate, 8 rows by 12 columns.
There is not a tremendous amount of data created in terms of bytes, but the plate map data file integrity is very important. Assay hits are elucidated based upon the plate maps. The Windows' versions of the Zymark machine interface software allow I/O functions with any available drive. For maximum reliability, it was decided to send all files generated to the local PC hard drive. To address the various causes of data loss, drive backups are a standard practice. Regular backups of the target hard drive produce data redundancy, but not one hundred percent. If there is a loss of the use of the hard drive, all files written since the last backup can only be rescued via a professional data recovery service.
One way to enable true fault tolerance is through the use of RAID in addition to a backup routine. In the situation described above, RAID level 1 (mirroring) was chosen as the solution. The simplest way to implement mirroring was to add a second hard drive and run the software RAID coded into Microsoft Windows NT™. Other software RAID products were examined, but all were abandoned. Software RAID packages have a heavy processor overhead. Combined with the fact that there was a Zymark controller connected to the PC, it was felt that reliable performance could not be realized. While researching possible vendors for a RAID hardware purchase, it was assumed that RAID was only available for the SCSI interface. There are actually a few IDE RAID manufacturers out there. In terms of cost, IDE RAID is cheaper than SCSI. The host adapters are much less expensive than their SCSI counterparts. Drives with the IDE interface are lower in price than SCSI hard drives.
Most RAID gear on the market relies on a combination of firmware and a software driver to operate. A supplier was found that sold an IDE RAID 1 card that did not use drivers. No drivers were required since the card only supported mirroring. By not requiring drivers, the card was compatible with any operating system. All necessary code was in the firmware. Further, the firmware would place minimal additional demands on CPU performance.
An Arco (www.arcoide.com) PCI IDE RAID adapter card was purchased. Two identical 1.0GB Quantum IDE hard disk drives were available as surplus items. First, the contents of the boot hard drive from the computer interfaced to the Zymark system were cloned onto one of the Quantum's using Symantec Ghost. Next it was a matter of mounting the two Quantum drives, inserting the card, then attaching the cables. Finally, using software provided by Arco, the mirror drive is built and initialized for use.
The Arco adapter does not allow hot swapping, but in the case of a failure, there will be an audible alarm. The surviving drive continues on, successfully receiving data from the controller for the remainder of the run. Once the run is completed, then the PC can be shut down and the bad drive replaced. To get back up and running, the new drive is made the mirror.
With RAID level 1, there are 2 copies of the hard drive, thus the chances of both drives failing at the same time are remote. In the case of single disk failure, there is no need to perform a restore of the replacement disk using the backups. Instead, the drive that did not go down is cloned onto the new drive, since it contains the identical information that the failed unit was storing.
Mirroring is the simplest of the RAID configurations, but also the least efficient. For example, mirroring results in greater overhead for disk ECC tasks. Another issue is the nature of the data being stored on the fault tolerant system. In the situation described above, many small writes occur, generating an I/O workload for the cpu in addition to servicing the rest of the system. The system performance hit due to the RAID card is minimized by invoking hardware implementation and not software. The Arco product does not support hot swapping of drives. The affected machine must be shut down before a drive can be replaced.
The section that follows shows a comparison of Quantum's performance ratings for the model Fireball 1080AT hard drives being used against data gathered from Winbench® 99, a widely available PC performance evaluation utility. The computer is a Dell Dimension XPS Pro200n, with 64 MB of RAM, BIOS revision A06.
Special Thanks to: Advanced Computer and Network Corporation (http://www.acnc.com/) for providing the RAID 1 diagram at http://www.acnc.com/04_01_01.html. ARCO Computer Products (http://www.arcoide.com/) for allowing use of the DupliDisk™ PCI card diagram from page 4 of the manual at http://www.arcoide.com/pdfs/pdfddpci.pdf.
Winbench results:
The CPU utilization was 4.11%.
Common test settings: Disk Drive=d:\\ Report CPU Utilization=Yes
Windows tasks were running during this test and could affect the test results
The CPU utilization was 2.35 %.
The transfer rate during this test averaged 4000 thousand bytes per second.
The playback directory was d: \\∼wbdtmp
The CPU Utilization was 20.37%.
Quantum Fireball 1080AT IDE Performance Specifications:
