DATA STORAGE ISSUES

CERN Storage Systems, Data Management, FATMEN, TMS, Unitree

It was noted that remote tape is troublesome because recovery by tape repositioning is hard to do, while disk is cheap and scalable so copying a while to disk seems much better than reading/writing tapes from an application. The idea is to have tape servers that respond over the net to any tape request. So far, such function's performance has been benchmarked at CERN at about 1.0-1.5 MBps, while about 800 KBps is actually seen using the "parallel channel" interface, probably because of the high level of interrupts required for such an interface.. The ESCON interface is faster though CERN had not yet measured it. They expect actual performance with ESCON to be 1.5-2.0 MBps rather than the current 800 KBps. CERN uses 32 KB blocks with FDDI on IBM RS6000s. They note that the ESCON interface requires a 5xx deskside model, more money that a 3xx desktop, and that the ESCON interface itself is quite expensive. They have lots of parallel channel tape drives and expect to move to the IBM NTP tape drives with Wide SCSI when they are available. They are rated at 9MBps. Their testing with a DEC Alpha doing disk to net gives about 6 MBps and expect tape to be a bit better.

The SHIFT software is currently available from anonymous FTP and is installed at places like DESY, FNL, and IN2P3.

TMS has proven to be useful. It is written as a SQL-DS application originally but has now been entirely rewritten in IBM C. It runs on a separate IBM 4381 but has been ported to Solaris 2.x and oracle 7 on a Sun and is awaiting testing and tuning. TMS will let you have write-protected tape volumes and will let FATMEN do dynamic volume allocation. Both DEC and IBM will have competing products; DEC's will be MLM and IBM's will probably be based on ADSM. Concerns were also expressed about the usefulness of NSL Unitree. It was pointed out that it cannot backup the metadata while the system is up, that a default maximum of 8000 tapes is in effect without changing the source, and that a recent review of such system on the net showed Unitree as a poor choice. CERN is working on new stager code. It will start a daemon to start the copy process and watch it. If the staging disk fills up before copying is done, it will suspend the copy, do disk space garbage collection, and then resume the copying. The staging software at CERN knows how to deal with SL tape.

IBM will be demoing the NTP technology at the IEEE MSS Symposium and probably at CHEP '94 in San Francisco. CERN expects it to be available in mid-1994. NTP has logical volume support. The pricing and connectivity of NTP will determine the market size. Competitors to IBM's technology are be STK and DEC with something called DLT. CERN is buying DLTs to replace 8mm tape backup systems. They're rated at 1.25 MBps and have been measured at 1.0 MBps.

CERN is curious about our STK experience and our future with D3 technology. They have one experiment that is going to collect data on a Sony DL21000 D1 system.

CERN Robotics, Exabytes, and General Operations

The PDP group now encompasses the robotics and tape vault operations with 10 shift operators plus 1 operator for consumables plus 25 contract operators to handle manual tape drives. CERN's involvement began with an IBM Joint Study and an acquisition in 1988. They expect to upgrade to the IBM 10 GB SCSI linear serpentine cartridge technology that is rated at 7-8 MBps. Right now about 50% of the tape mounts are manual. Tape data is staged to disk and erase later when disk space is needed. Right now the robotics are controlled by VM but they have just received the RS6000 software. The current bottleneck is in the control unit, capable of a total of 6 MBps throughput even though each drive can do 6 MBps.

CERN notes that using 8mm Exabyte tapes is taking a step backwards in reliability and performance. It is really only useful for backups on standalone systems where the tapes will be read again on the same tape drive or for transporting data to other places that have inexpensive tape drives. They have a user-operated station for copying 3480 cartridges to/from 8mm tape. The use an 8500 drive, because it is more reliable than the 8200, without compression. They use two DECstation for self-service. For robotics, they also have 8mm tape drives attached to an IBM channel via a SCSI-channel converter. They find that 8mm tape drive heads wear out about every 2 month and so they expect to replace tape drives on a regular basis. The had tried the Summus tape carousel and found it was a poor choice; the Exabyte carousel worked well and it was easier to replace a tape drive yourself when it became necessary.

DESY Data Management

DESY is interested in the Lachmann software, rather than Unitree. They want a central data repository that conforms to the IEEE Mass Storage Model. DESY expects to use a Gigarouter from NetStar to connect various media together (HIPPI, FDDI, etc.). Phase 1 now has the IBM ES9000 as the silo controller. Phase 2 will disconnect one silo from the IBM ES9000. They are shopping for a sophisticated Hierarchical Storage Manager. This is an HSM that need to be able to intelligently choose between STK and Ampex. They feel Ampex will be good for sequentially accessing large data sets with a faster search, though they only have 6 drives. STK will be better for access to smaller amounts, with 36 drives.

DESY has been very happy with their SGI machines. The SGI Challenge has a 1.2 GBps bus, has IO processors that can do 320 MBps, and can have 32 SCSI busses off the IOPs without going through the VME bus. If striped, they can get a read rate > 11 MBps and a write rate > 7 MBps.

Martin Gasthuber discussed hierarchical storage management. DESY has chosen OSM (Open Storage Manager) from Lachman. The concept is that data is produced with intelligent controllers and a central "bitfile" server. The HSM discovers the most recent copy of your data, talks to the Storage Server, and then a direct communication occurs between the Storage Server and the client making the request. The client would have a Migration Filesystem on top of a standard filesystem to make secondary storage appear primary. The Migration Filesystem is the typical client of the Storage Server. OSM clients can be NFS, AFS database access, or even Fatmen. DESY has a license and the package has arrived. They noted that IBM Adstar has also licensed OSM.

Michael Ernst then discussed their Ampex tape system. DESY is quite happy with the system and probably will not buy any further STK systems. With new software, they expect to do better than 14 MBps on reads and writes. One question that always comes up is tape wear, tape re-readability, and head wear and replacement. Michael said that a head appears to be good for about 1000 tape/head contact hours. This is equivalent to reading or writing about 20-30 TB of data. They have also test readability of tapes and found that they could reread tape more than 1500 times with no problems. When the bit error rate begins to climb, they clean the heads. If error rates are still a problem, then they change the head assembly (8 heads). This turns out to take about 15 minutes and is self-service -- no Ampex technician needed. Such a head assembly costs about $2500.

RAL Storage Management

RAL has developed a data storage with migration facility. Currently it is based on an IBM 3090-400 running VM which incorporates STK Silos with over 100 GB of disk staging space. On VM they modified the LINK command to load the data from tape to disk space if necessary. Much of the impetus was to minimize the number of drives required in the STK Silos. No user tapes are contained in the Silos. Data is moved from the Silos to DAT tapes when space is required in the Silos. Network access to the Silos is via 3 RS6000's with channel interfaces connected to the VM system.

RAL looked at various tape technologies for archive purposes. They had a Metrum SVHS 1/2" 6 TB Robot on site for evaluation but were concerned about reliability, expense single vendor support and rejected it. They compared DAT and 8MM. The costs were about equal (though the media costs were greater than for SVHS). They Chose DAT since it turned out to be much more reliable and bought 12 DAT drives. These are operator serviced with one mount/drive every 3 hours (based on data rates). They looked at stackers to possibly reduce operator intervention however it was not attractive costwise compared to simply increasing the number of drives. RAL has one 8 mm drive for compatibility purposes. The DAT data rates are (slower than 8 mm) 183 KB/sec today, will increase to 366 KB sec and are soon expected to go to 510 KB/sec. The DAT tapes are 90 meters long and hold about 2 GB/tape. RAL is copying about 30K 3420 type tapes to DAT tapes. The reason for this is to get rid of the older disintegrating tapes, to enable the data to be accessable to current devices, to reduce the floor space for storage needed by probably two orders of magnitude, and to record the data in a well defined DAT format to make future retrieval easier.

RAL are looking at the new IBM Digital Linear tapes, which look attractive when compared to STK. Particular STK concerns are maintenance costs, uncertainty about their ability to deliver the helical scan drives, and the reliability of the helical scan drives.