(C) Copyright Greg Smith, 2002-2007
Shared device support allows multiple Hercules instances to share devices. The device will be 'local' to one instance and 'remote' to all other instances. The local instance is the 'server' for that device and the remote instance is the 'client'. You do not have to IPL an operating system on the device server. Any number of Hercules instances can act as a server in a "Hercplex".
To use a device on a remote system, instead of specifying a file name on the device config statement, you specify
0100 3350 localhost:3990:0100
which says there is a device server on the local host listening on port 3990 and we want to use its 0100 device as 0100. The default port is 3990 and the default remote device number is the local device number. So we could say:
0100 3350 localhost
instead, providing we don't actually have a file 'localhost'. Interestingly, the instance on the local host listening on 3990 could have a statement:
0100 3350 192.168.200.1::0200
which means that instance in turn will use device 0200 on the server at 192.168.200.1 listening on port 3990. The original instance will have to 'hop' thru the second instance to get to the real device.
Device sharing can be 'split' between multiple instances. For example, suppose instance A has:
SHRDPORT 3990 0100 3350 localhost:3991 0101 3350 mvscat
and instance B has:
SHRDPORT 3991 0100 3350 mvsres 0101 3350 localhost
Then each instance acts as both a client and as a server.
When 'SHRDPORT' is specified, thread 'shared_server' is started at the end of Hercules initialization. In the example above, neither Hercules instance can initialize their devices until the server is started on each system. In this case, the device trying to access a server gets the 'connecting' bit set on in the DEVBLK and the device still needs to initialize. After the shared server is started, a thread is attached for each device that is connecting to complete the connection (which is the device init handler).
There are (at least) two approaches to sharing devices. One is to execute the channel program on the server system. The server will need to request from the client system information such as the ccw and the data to be written, and will need to send to the client data that has been read and status information. The second is to execute the channel program on the client system. Here the client system makes requests to the server system to read and write data.
The second approach is currently implemented. The first approach arguably emulates 'more correctly'. However, an advantage of the implemented approach is that it is easier because only the client sends requests and only the server sends responses.
Both client and server have a DEVBLK structure for the device. Absurdly, perhaps, in originally designing an implementation for shared devices it was not clear what type of process should be the server. It was a quantum leap forward to realize that it could just be another hercules instance.
(If this section is as boring for you to read as it was for me to write then please skip to the next section ;-)
The client sends an 8 byte request header and maybe some data:
+-----+-----+-----+-----+-----+-----+-----+-----+ | cmd |flag | devnum | id | length | +-----+-----+-----+-----+-----+-----+-----+-----+ <-------- length ---------> +----- . . . . . -----+ | data | +----- . . . . . -----+
'cmd' identifies the client request. The requests are:
NOTE: This action should actually be SETOPT or some such; it was just easier to code a COMPRESS specific SETOPT (less code).
'flag' qualifies the client request and varies by the request.
'devnum' identifies the device by number on the server instance. The device number may be different than the device number on the client instance.
'id' identifies the client to the server. Each client has a unique positive (non-zero) identifier. For the initial CONNECT request 'id' is zero. After a successful CONNECT, the server returns in the response header the identifier to be used for all other requests (including subsequent CONNECT requests). This is saved in dev->rmtid.
'length' specifies the length of the data following the request header. Currently length is non-zero for READ/WRITE requests.
The server sends an 8 byte response header and maybe some data:
+-----+-----+-----+-----+-----+-----+-----+-----+ |code |stat | devnum | id | length | +-----+-----+-----+-----+-----+-----+-----+-----+ <-------- length ---------> +----- . . . . . -----+ | data | +----- . . . . . -----+
'code' indicates the response to the request. OK (0x00) indicates success however other codes also indicate success but qualified in some manner:
'stat' contains status information as a result of the request. For READ/WRITE requests this contains the 'unitstat' information if an IOERR occurred.
'devnum' specifies the server device number
'id' specifies the system identifier for the request.
'length' is the size of the data returned.
Cached records (eg CKD tracks or FBA blocks) are kept independently on both the client and server sides. Whenever the client issues a START request to initiate a channel program the server will return a list of records to purge from the client's cache that have been updated by other clients since the last START request. If the list is too large the server will indicate that the client should purge all records for the device.
Data that would normally be transferred uncompressed between client
and host can optionally be compressed by specifying the '
keyword on the device configuration statement or attach command.
0100 3350 192.168.2.12 comp=3
The value of the 'comp=' keyword is the zlib compression parameter which should be a number between 1 .. 9. A value closer to 1 means less compression but less processor time to perform the compression. A value closer to 9 means the data is compressed more but more processor time is required.
If the server is on 'localhost' then you should not specify 'comp='. Otherwise you are just stealing processor time to do compression/ uncompression from hercules. If the server is on a local network then I would recommend specifying a low value such as 1, 2 or 3. We are on a curve here, trying to trade cpu cycles for network traffic to derive an optimal throughput.
If the devices on the server are compressed devices (eg CCKD or CFBA) then the 'records' (eg. track images or block groups) may be transferred compressed regardless of the 'comp=' setting. This depends on whether the client supports the compression type (zlib or bzip2) of the record on the server and whether the record is actually compressed in the server cache.
Suppose on the client that you execute one or more channel programs to read a record on a ckd track, update a record on the same track, and then read another (or the same) record on the track.
For the first read the server will read the track image and pass it to the client as it was originally compressed in the file. To update a portion of the track image the server must uncompress the track image so data in it can be updated. When the client next reads from the track image, the track image is uncompressed.
Specifying 'comp=' means that uncompressed data sent to the client will be compressed. If the data to be sent to the client is already compressed then the data is sent as is, unless the client has indicated that it does not support that compression algorithm.
Greg Smith firstname.lastname@example.org
Last updated $Date$ $Revision$