4 ms·

Could you elaborate on how it did things better than the common cluster file systems of today? (OCFS2, Hadoop, Ceph, etc.)

by wogna 9y ago

valarauca1 9y ago
Okay so OpenVMS wasn't just FileSystem clustering. It was physical PC clustering. So I group of networked computers appeared like 1 physical computer. While a MESO or DCOS can do this today, they don't offer durability. On OpenVMS clusters if the machine running a job physically explodes you don't lose data. RAM and process state is network synchronized. This allows for rolling upgrades. Your cluster can update and reboot 1 box at a time, but your apps never actually stop, TCP connections are never dropped ETC. For large corporations who've gotten used to these features they're awesome to hold onto.
- reitanqild 9y ago
  TCP connections are never dropped ETC One thing I wonder now: how can this work? Is there a physical load balancer in front or something?
  - eej71 9y ago
    That claim is simply not true.
- MrBuddyCasino 9y ago
  "RAM and process state is network synchronized." Is this as slow as it sounds?
  - staofbur 9y ago
    Yes. I saw the last days of a VMS cluster (actual VAX not Alpha) system being ported to windows/vb/com/asp. The speed difference was actually incredible. It ended up on a mediocre single Xeon 500MHz P3. At the end of the day, the windows based system was just as reliable as far as nines went and tens of times faster. This was only just recently replaced with asp.net I understand. They got nearly 20 years out of each rewrite.
    - cat199 9y ago
      You're confusing hardware speed with the software/IPC measurement - the fastest VAX machine ever made ran at 100Mhz (or maybe 120), with 10Mbit ethernet and SCSI-2 I/O paths (1MB/s on a disk or so). If you found a 486/100 and compared raw compute speed, the vax would likely have won for most cases, because it has 4x the registers. I'm not actually sure about the clustering IPC performance, but you are essentially comparing hardware about 3-4 generations apart and blaming the software for the issues.. if you were comparing with 500MHz Alphas, etc, then you might have a point..
      - staofbur 9y ago
        Not quite. We scaled it to 50x the user count at the same time.
srean 9y ago
Dont know about OCFS2 but Hadoop's file system HDFS and Ceph are hardly the stellar examples of rock solid and performant (I know its not an English word) distributed file systems. Google's mapreduce was so much more efficient in speed (4X and 4x last time I played with those tools) and resources primarily because of the bottlenecks imposed by HDFS design and implementation.
- garettmd 9y ago
  If performant isn't actually a word than a lot of people (myself included) have been doing it wrong. I can't think of a different variation on the word performance that would work
  - srean 9y ago
    Some people are allergic to it, as for myself i cannot find a better substitute.
voidmain0001 9y ago
I don't have VMS experience, but rather AS/400 and iSeries experience. The difference vs generic purpose OS (Linux, Windows, OSX) is that the AS/400, and I'm going to assume VMS since they operated in similar markets, had these features baked in by the manufacturer. OS/400 on the AS/400 was a menu based system. If the menu option existed, the feature was available, and it was as simple as flipping the bit to YES to enable a feature. Enable clustering? Select Yes, add the information of the cluster and you were done. It just worked. Did it work well? I guess. There was no way to know if it could be done better since it was a manufacturer's enabled feature. Of course Linux, Windows, OSX can do the same thing as OS/400 on AS/400 but with a lot of configuration, and a lot of jumping through hoops [it's not called Hadoop for nothing ;-)].
pedasmith 9y ago
For example: you could make a cluster-wide lock that you code uses to coordinate access to cluster-wide shared resource. Locks on most OSes have an "outer scope" of a process or the machine.

Could you elaborate on how it did things better than the common cluster file systems of today? (OCFS2, Hadoop, Ceph, etc.)

by wogna 9y ago

Could you elaborate on how it did things better than the common cluster file systems of today? (OCFS2, Hadoop, Ceph, etc.)

valarauca1 9y ago

Okay so OpenVMS wasn't just FileSystem clustering. It was physical PC clustering. So I group of networked computers appeared like 1 physical computer. While a MESO or DCOS can do this today, they don't offer durability. On OpenVMS clusters if the machine running a job physically explodes you don't lose data. RAM and process state is network synchronized. This allows for rolling upgrades. Your cluster can update and reboot 1 box at a time, but your apps never actually stop, TCP connections are never dropped ETC. For large corporations who've gotten used to these features they're awesome to hold onto.

reitanqild 9y ago
TCP connections are never dropped ETC One thing I wonder now: how can this work? Is there a physical load balancer in front or something?
- eej71 9y ago
  That claim is simply not true.
MrBuddyCasino 9y ago
"RAM and process state is network synchronized." Is this as slow as it sounds?
- staofbur 9y ago
  Yes. I saw the last days of a VMS cluster (actual VAX not Alpha) system being ported to windows/vb/com/asp. The speed difference was actually incredible. It ended up on a mediocre single Xeon 500MHz P3. At the end of the day, the windows based system was just as reliable as far as nines went and tens of times faster. This was only just recently replaced with asp.net I understand. They got nearly 20 years out of each rewrite.
  - cat199 9y ago
    You're confusing hardware speed with the software/IPC measurement - the fastest VAX machine ever made ran at 100Mhz (or maybe 120), with 10Mbit ethernet and SCSI-2 I/O paths (1MB/s on a disk or so). If you found a 486/100 and compared raw compute speed, the vax would likely have won for most cases, because it has 4x the registers. I'm not actually sure about the clustering IPC performance, but you are essentially comparing hardware about 3-4 generations apart and blaming the software for the issues.. if you were comparing with 500MHz Alphas, etc, then you might have a point..
    - staofbur 9y ago
      Not quite. We scaled it to 50x the user count at the same time.

srean 9y ago

Dont know about OCFS2 but Hadoop's file system HDFS and Ceph are hardly the stellar examples of rock solid and performant (I know its not an English word) distributed file systems. Google's mapreduce was so much more efficient in speed (4X and 4x last time I played with those tools) and resources primarily because of the bottlenecks imposed by HDFS design and implementation.

garettmd 9y ago
If performant isn't actually a word than a lot of people (myself included) have been doing it wrong. I can't think of a different variation on the word performance that would work
- srean 9y ago
  Some people are allergic to it, as for myself i cannot find a better substitute.

voidmain0001 9y ago

I don't have VMS experience, but rather AS/400 and iSeries experience. The difference vs generic purpose OS (Linux, Windows, OSX) is that the AS/400, and I'm going to assume VMS since they operated in similar markets, had these features baked in by the manufacturer. OS/400 on the AS/400 was a menu based system. If the menu option existed, the feature was available, and it was as simple as flipping the bit to YES to enable a feature. Enable clustering? Select Yes, add the information of the cluster and you were done. It just worked. Did it work well? I guess. There was no way to know if it could be done better since it was a manufacturer's enabled feature. Of course Linux, Windows, OSX can do the same thing as OS/400 on AS/400 but with a lot of configuration, and a lot of jumping through hoops [it's not called Hadoop for nothing ;-)].

pedasmith 9y ago

For example: you could make a cluster-wide lock that you code uses to coordinate access to cluster-wide shared resource. Locks on most OSes have an "outer scope" of a process or the machine.