On 2022-06-28 21:15, Honnappa Nagarahalli wrote:
<snip>
From: Mattias Rönnblom [mailto:mattias.ronnb...@ericsson.com]
Sent: Monday, 27 June 2022 13.06
Hi.
Is it safe to enable stats on MT safe services?
https://protect2.fireeye.com/v1/url?k=31323334-501d5122-
313273af
-
4
54445555731-6096fdb16385f88f&q=1&e=27b94605-d1e2-40b6-
af6d-
9ebc54d
5db18&u=https%3A%2F%2Fgithub.com%2FDPDK%2Fdpdk%2Fblob%2Fmain%
2Flib
%2Feal%2Fcommon%2Frte_service.c%23
L3
6
6
It seems to me this would have to be an __atomic_add for this
code to produce deterministic results.
I agree. The same goes for the 'calls' field.
The calling function does the locking.
https://protect2.fireeye.com/v1/url?k=31323334-501d5122-
313273af
-
454
445555731-5ce419f8bf9f9b76&q=1&e=27b94605-d1e2-40b6-af6d-
9ebc54d5db1
8&u=https%3A%2F%2Fgithub.com%2FDPDK%2Fdpdk%2Fblob%2Fmain%2Flib
%2Feal
%2Fcommon%2Frte_service.c%23L3
98
For more information you can look at:
https://protect2.fireeye.com/v1/url?k=31323334-501d5122-
313273af
-
454
445555731-ba0d1416f08856f0&q=1&e=27b94605-d1e2-40b6-
af6d-
9ebc54d5db1
8&u=https%3A%2F%2Fgithub.com%2FDPDK%2Fdpdk%2Fblob%2Fmain%2Flib
%2Feal
%2Finclude%2Frte_service.h%23L
120
What about the
https://protect2.fireeye.com/v1/url?k=31323334-501d5122-
313273af-
4544
45555731-b64334addc78c264&q=1&e=27b94605-d1e2-40b6-af6d-
9ebc54d5db18&
u=https%3A%2F%2Fgithub.com%2FDPDK%2Fdpdk%2Fblob%2Fmain%2Flib%2
Feal%2F
common%2Frte_service.c%23L404
call (for MT safe services)?
There's no lock held there.
Good point.
This is the case where the service running in service cores is MT
safe. However,
the stats are incremented outside of the MT Safety mechanism
employed by the service. So, yes, this and other updates in the
function 'service_runner_do_callback' need to be updated atomically.
Maybe a better solution would be to move this to the core_state
struct (and eliminate the "calls" field since the same information
is already in the core_state struct). The contention on these cache
lines will be pretty crazy for services with short run time (say a
thousand cycles or less per call), assuming they are mapped to many
cores.
That's one option, the structures are internal as well. With this
option stats
need to be aggregated which will not give an accurate view. But, that
is the nature of the statistics.
Per-core counters is a very common pattern. Used for Linux MIB
counters, for example. I'm not sure I think it's much less accurate.
I mean, you just load in quick succession what's globally visible for
the different per-lcore counters. If you do a relaxed store on an
ARM, how long time does it take until it's seen by someone doing a relaxed
load on a different core? Roughly.
I think my explanation of the problem is not clear.
If a service is running on more than one core and the stats are per core, when
we aggregate, the resulting statistics is not atomic. By making the stats per
core,
we will be taking out that feature which is present currently (even though it is
implemented incorrectly). As we agree, the proposed change is a common
pattern and taking away the atomicity of the stats might not be a problem.
Isn't it just a push model, versus a pull one? Both give just an approximation,
albeit a very good one, of how many cycles are spent "now" for a particular
service. Isn't time a local phenomena in a SMP system, and there is no global
"now"? Maybe you can achieve such with a transaction or handshake of some
sort, but I don't see how the an __atomic_add would be enough.
If we consider a global time line of events, using atomic operation will
provide a single 'now' from the reader's perspective (of course there might be
writers waiting to update). Without the atomic operations, there will not be a
single 'now' from reader's perspective, there will be multiple read events on
the timeline.
At the time of the read operation (in the global counter solution),
there may well be cycles consumed or calls having been made, but not yet
posted. The window between call having been made, and global counter
having been incremented (and thus made globally visible) is small, but
non-zero.
(The core-local counter solution also use atomic operations, although
not __atomic_add, but store, for the producer, and load, for the consumer.)
I was fortunate to get some data from a real-world application, and enabling
service core stats resulted in a 7% degradation of overall system capacity. I'm
guessing atomic instructions would not make things better.
Is the service running on multiple cores?
Yes. I think something like 6 cores were used in this case. The effect
will grow with core count, obviously. On a large system, I don't think
you will do much else but fight for this cache line.
If you want post counter updates to some shared data structure, you need
to batch the updates to achieve reasonable efficiency. That will be,
obviously, at the cost of accuracy, since there will be a significant
delay between local-counter-increment, and
post-in-global-data-structure. The system will be much less able to
answer how many cycles have been consumed at a particular point in time.
For really large counter sets, the above, batched-update approach may be
required. You simply can't afford the memory required to duplicate the
counter struct across all cores in the system. In my experience, this
still can be made to meet real-world counter accuracy requirement.
(Accuracy in the time dimension.)
I am also wondering if these stats are of any use other than for debugging.
Adding a capability to disable stats might help as well.
They could be used as a crude tool to determine service core utilization.
Comparing utilization between different services running on the same core
should be straight-forward, but lcore utilization is harder in absolute terms.
If
you just look at "cycles", a completely idle core would look like it's very busy
(basically rdtsc latency added for every loop). I assume you'd have to do
some
heuristic based on both "calls" and "cycles" to get an estimate.
I think service core utilization would be very useful, although that would
require
some changes in the service function signature, so the service can report
back if
it did some useful work for a particular call.
That would make for a DPDK 'top'. Just like 'top', it can't impose any serious
performance degradation when used, to be really useful, I think.
Sure, it should be possible to turn it on and off. I thought that was the case
already?
Thanks, yes, this exists already. Though the 'loops' counter is out of the stats
enable check, looks like it is considered as an attribute for some reason.
Idle service cores will basically do nothing else than stall waiting
for these lines, I suspect, hampering the progress of more busy cores.
