| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 314.1 | | COOKIE::FROEHLIN | Let's RAID the Internet! | Tue Jan 28 1997 09:26 | 29 |
| Marc,
>1) Very low transfer rates - much lower than when they
> read/write to their personal share on the same system;
To exclude any non-RAID issues, could you copy files to/from the RAID
set and compare how long it takes to copy these files to/from a regular
disk?
"their personal share on the same system" means what?
RAID 5 is not a good candidate for write operations. In typical cases
you get 4 times more I/Os for a write than on a normal disk.
>2) Random transmission errors, which result in corrupted
> files after a copy, in both directions.
Can you do an "ANALYZE/DISK/READ DPAnnn:"? I would expect files with
parity errors. If not, I need more details on the symptoms of the
errors.
>So my question is : is PATHWORKS compatible with the VMS
>RAID software ? Has this been explicitely tested ?
Since DPA devices by all means behave like regular disk devices there
shouldn't be any incompatibilities here. There shouldn't...but we have
not explicitly tested Pathworks together with RAID Software.
Guenther
|
| 314.2 | File corruption=false alarm;still performance problems. | TPLAB::VANDYCK | Symbolic stack dump fellow | Tue Feb 11 1997 07:19 | 21 |
| I have asked the users that reported the problems to
me to make some additional tests.
It seems that they have been able to reproduce the
file corruption problem on another disk (i.e. not
raid) so the RAID software has nothing to do with this.
We'll take that with Pathworks. Please accept my
apologies for the false alarm.
On the other hand, we still have the performance
problem that is worrying us. Yes, writes to RAIDsets
generate more IOs than to single disks, but with a
RAIDset, those IOs are distributed on n disks rather
than on one only, right ? So in elapsed time, it should
be the same, if not faster ?
The experiences we made reported that the raid set
is approximately 3 times slower, in average, than
a single disk. There again, I will ask for some
additional tests and report the results when they are
available.
|
| 314.3 | | COOKIE::FROEHLIN | Let's RAID the Internet! | Tue Feb 11 1997 09:14 | 21 |
| .2>We'll take that with Pathworks. Please accept my
.2>apologies for the false alarm.
Accepted :-)
.2>problem that is worrying us. Yes, writes to RAIDsets
.2>generate more IOs than to single disks, but with a
.2>RAIDset, those IOs are distributed on n disks rather
Not quite! On RAID 5 typically with small I/Os (<chunk size)
first the parity blocks are read and the old data blocks are read which
can be done in parallel. Next the old data is backed out of the
parity in memory and the new parity and data is written back to disk
which again can be done in parallel. If the RAID set member disks are
connected to the same bus/controller/interconnect some serialization
has to happen which means less parallelism.
What is their system configuration around the RAID sets?
Thanks
Guenther
|
| 314.4 | Detail config | TPLAB::VANDYCK | Symbolic stack dump fellow | Wed Feb 12 1997 03:36 | 18 |
| Detailed configuration :
- Cluster of two VAX 6610 running VMS V6.1 ;
- Raid set made of 6 RA 90 disks ;
- The six disks are all dual-ported between two
HSC 95 controllers;
- The 6 disks are connected to 4 different requestors.
- The single disks used to make the comparison are
RA 73 disks, also dual-ported between the same
controllers.
BTW, this is an internal system, not a customer one.
It means you can have access to it if you whish so.
|