DPL: improve resource management#2513
Conversation
ktf
changed the title
ktf
force-pushed
the
improved-resource-allocation
branch
2 times, most recently
from
7c0726b to
5d0a219
Compare
ktf
changed the title
|
@aalkin this now seems to compile and work. This is meant to pave the way for shared memory usage (via DPL driver) and static multiple-host support. |
ktf
force-pushed
the
improved-resource-allocation
branch
from
5d0a219 to
ed6fde6
Compare
|
@ktf I've tested in heaviest reco flows, they are not affected. |
|
@ktf it seems the test_DeviceSpec and test_Graphviz both fail due to random int in place of a port number, most likely because we are not setting port and range in DefaultOffer and ComputingOffer struct itself has no default initialization. |
Rework the way DPL assign resources to devices. This is a first step in moving away from the hardcoded 127.0.0.1 ip address and allows DPL workflows to run distributed without a third party configuring the channels. A further optimization will also allow DPL to configure ipc / shared memory when two devices happen to be assigned to the same machine.
ktf
force-pushed
the
improved-resource-allocation
branch
from
ed6fde6 to
40a79c1
Compare
|
QC seems to be doing fine. |
|
Ok. I am merging this and I open a new one which optimises local connections to "ipc" (i.e. unix sockets). We can then try with "inproc" (i.e. managed shared memory) but we cannot get it to work on my mac, yet. |
|
Sorry, I meant |
Rework the way DPL assign resources to devices. This is a first step in moving away from the hardcoded 127.0.0.1 ip address and allows DPL workflows to run distributed without a third party configuring the channels. A further optimization will also allow DPL to configure ipc / shared memory when two devices happen to be assigned to the same machine.
Rework the way DPL assign resources to devices. This is a first step in moving away from the hardcoded 127.0.0.1 ip address and allows DPL workflows to run distributed without a third party configuring the channels. A further optimization will also allow DPL to configure ipc / shared memory when two devices happen to be assigned to the same machine.
4 participants
Rework the way DPL assign resources to devices. This is a first
step in moving away from the hardcoded 127.0.0.1 ip address and
allows DPL workflows to run distributed without a third party
configuring the channels.
A further optimization will also allow DPL to configure ipc / shared memory when two devices happen to be assigned to the same machine.