brodie@fps.mcw.edu (01/10/91)
The following info is direct from my local DEC office. Since
I am going from a simple config to dual HSC's and full volume
shadowing, I needed some guidelines. Here they are.
Enjoy!
HSCXX configuration and performance considerations pg=13
SUBJECT: PERFORMANCE CONSIDERATIONS FOR HSC CONTROLLERS.
27-MARCH-1989
1 INTRODUCTION
There have always been performance considerations with regards to
configuring devices on the HSC. Until recently, these considerations
were important, but not extremely critical because the speed of the
available devices did not tax the bandwidth of the HSC. With the
introduction of the RA82, RA90, TA90, and RA70, configuring for
maximum performance has become very important. In addition, if this
is not done, not only will there be a effect in performance, but data
buss contention caused errors may also occur, such as data buss
overruns, drat seek timeouts, and EDC errors.
Some general rules can be given, but things such as the
customer's application and the usage of the devices also need to be
considered when looking at the overall configuration. This memo will
try to provide some general guidelines and some understanding of what
effects performance in the HSC and and to what degree.
There are several things to keep in mind when configuring a
device on a HSC.
1. Only one device on a requestor can transfer data at one time.
2. Shadow set members need to be on separate requestors.
3. Requestor priority.
4. The speed of the device.
Sections 2, 3 and 4 will discuss these points in detail. Section 5
will offer some configuration guidelines, and section 6 discusses a
performance issue specific to single ported drives.
2 ONLY ONE DEVICE ON A REQUESTOR CAN TRANSFER DATA AT A TIME.
Each requestor can support 4 devices, however, only one of these
devices can transfer data at any one time. While one device is
transferring data, other devices can be seeking or positioning tape.
This means, the fewer fast and heavily used devices on any one
requestor the better the performance.
There is no significant performance difference between individual
ports within a requestor.
3 SHADOW SET MEMBERS MUST BE ON SEPARATE REQUESTORS.
This is a most important performance consideration. When the HSC
picks which drive to read from on a shadow set, the first check made
is for members on the same requestors. If all members are on the same
requestor, the HSC just uses the primary member and does no read
optimization at all. This means the same drive is always used on
reads and the speed advantage of reads in a shadow set are lost. On
writes, only one drive can be written to at any one time while
simultaneous writes could occur if the drives were split across
different requestors.
4 REQUESTOR PRIORITY AND DEVICE SPEED.
Requestor priority in the HSC is used two different ways. The
first, and the most obvious, is if two requestors are contending for
the data buss at the same time, the higher priority (higher numbered)
requestor wins.
The second way requestor priority is used is not so obvious and
requires some further explanation.
Each device connected to the HSC reports back it's transfer rate
to the HSC when it is brought on line. The HSC uses this to allocate
bandwidth on the data buss along with requestor priority. A RA90
needs every 3rd data buss cycle with it's high transfer speed so this
means 3 RA90s can transfer data at the same time. The RA70 needs only
every 5th cycle so 5 RA70s can transfer at the same time. This is
assuming the drives are on separate requestors. The requestor module
buffers less than 2 words, so if the RA90 cannot get the buss for 3
consecutive cycles, a data buss overrun will occur.
This works very well as long as the faster devices are at the
higher requestor numbers. If a RA90 is running with 3 RA70s, the RA90
will always get the buss when it needs it as long as it is at a higher
requestor number. However, if the RA90 is installed at a lower
requestor number than the RA70s, it is possible that 3 RA70s will lock
out the RA90 for 3 cycles an cause a overrun condition. To prevent
this, the HSC checks the speed and priority of the devices, and if it
detects a slower device at a higher priority, it allocates the slower
device the same bandwidth as the faster device.
For example, if RA70s were placed higher than RA90s, the HSC
would consider the RA70 the same speed as RA90s and would allow only 3
RA70s to transfer at the same time or any combination of 3
RA90s/RA70s. This ensures the RA90 will get at least the one out of
every 3 cycles it needs. This also means 40% of the HSC bandwidth
would not be utilized. For this reason, it is very important to
configure devices according to their relative transfer speeds. The
relative speeds of currently supported devices from the fastest to the
slowest is:
TA90
RA90
RA82
RA81
RA60
RA70
RA80
TA78/79/81
TA78/79/81 tape drives, unlike disk drives, do not have steady
transfer rates due to the irregularities of tape movement. Instead,
their transfer rates are very erratic and bursty in nature. For this
reason, they can cause bandwidth problems for other devices if given
higher priority on the data buss, even though their "average" transfer
rate is much slower. For this reason it is especially important to
keep them at the lowest requestor numbers. This means they cannot be
mixed on the same requestor with the cached TA90 which must be at the
highest requestor numbers.
5 GUIDELINES
Based on the above information, some basic guidelines can be
formulated to aid in configuring the HSC and it's devices. First,
configure the devices of different device types according to the
following criteria in order of priority;
1. TA78/79/81 tape drives must be at the lowest requestor
numbers. TA90s must be at the highest requestor numbers and
cannot share a requestor with TA78/79/81 type drives.
2. Shadow set members must be on separate requestors. This has
the highest performance impact and the highest priority.
3. Faster devices should be installed on higher requestor
numbers. This ensures the maximum possible transfer rate
will be maintained and eliminates the possibility of data
buss contention caused errors. Having shadow members on
separate requestors may conflict with this rule, in which
case, shadow members on different requestors takes priority.
4. Faster and high usage devices should be on requestors by
themselves, or as few of these devices on any one requestor
as possible.
Finally, after configuring the different device types using the above
guidelines, look at the devices of the same type. When configuring
devices of the same type, usage of each device needs to be considered
along with requestor priority and the fact that a requestor can only
transfer from one port at a time to determine the best configuration.
For example, the heaviest used device of one type, might be
placed on a requestor by itself, or with a device that is seldom used
and possibly placed at the highest requestor number of the devices of
that same type. Some judgment will have to be used here along with
some knowledge of the customer's application.
6 ADDITIONAL PERFORMANCE CONSIDERATION FOR SINGLE PORTED DISK DRIVES.
VMS version 5.0 increased the frequency of performing it's
Determine Access Path processing (DAP). DAP processing is used by VMS
to periodically cause the disk drives to release their ports so VMS
can determine all access paths to all drives. What this translates to
in the HSC/drive is a topology command being sent to the drive. This
topology command causes the drive to release it's selected port and
send an attention/available to the other port if the port button is
selected. During this time the drive is unavailable for disk
transfers, however, in a normal dual ported configuration it takes
only milliseconds for the attention/available and the HSC response.
The problem occurs when the drive has both port buttons selected
and is only connected to one controller. In that case, the drive
sends the attention/available on the unconnected port, and sits there
and waits for a response until it eventually times out after about 2
seconds. This means all IO to the drive stalls for 2 seconds every
time VMS calls DAP processing. Currently this will happen only every
6 minuets but VMS is expected to increase this frequency in the
future.
Even though DAP processing is relatively infrequent now, it is
important for the Customer and Field Service to understand that there
still is some effect on performance by keeping both port buttons
selected on single ported drives and this may be of larger concern in
the future.
//HSC50 HSC70 HSC40 HSC CONFIGURATION