RE: [PATCH 0/5] support telemetry dump dev

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> Sent: Thursday, 16 June 2022 10.19
> 
> On Wed, Jun 15, 2022 at 10:15:27PM +0200, Morten Brørup wrote:
> > > From: Chengwen Feng [mailto:fengchengwen@huawei.com]
> > > Sent: Wednesday, 15 June 2022 09.39
> > >
> > > This patchset contains five patch which support telemetry dump
> > > dmadev/eventdev/rawdev/ethdev.
> > >
> > > Chengwen Feng (5):
> > >   usertools: use non-strict when json-loads in telemetry
> > >   dmadev: support telemetry dump dmadev
> > >   eventdev: support telemetry dump eventdev
> > >   rawdev: support telemetry dump rawdev
> > >   ethdev: support telemetry private dump
> > >
> > >  lib/dmadev/rte_dmadev.c     | 39 +++++++++++++++++++++++++++++++++
> > >  lib/ethdev/rte_ethdev.c     | 42
> ++++++++++++++++++++++++++++++++++++
> > >  lib/eventdev/rte_eventdev.c | 43
> +++++++++++++++++++++++++++++++++++++
> > >  lib/rawdev/rte_rawdev.c     | 42
> ++++++++++++++++++++++++++++++++++++
> > >  usertools/dpdk-telemetry.py |  2 +-
> > >  5 files changed, 167 insertions(+), 1 deletion(-)
> > >
> > > --
> > > 2.33.0
> > >
> >
> > While hoping for someone to fix the Telemetry library bug regarding
> the missing JSON encoding of strings
> (https://bugs.dpdk.org/show_bug.cgi?id=1037), thus making the patch
> regarding usertools/dpdk-telemetry.py superfluous,
> >
> > Series-Acked-by: Morten Brørup <mb@smartsharesystems.com>
> >
> 
> I disagree with taking this series without fixing the underlying issue.
> I
> don't think we should be sending invalid json data so doing proper
> escaping of \n\t etc. characters on encode is the best solution.

OK. I'll accept upgrading from hope to requirement. :-)

> 
> There is also some merit in adding in support for tagging strings that
> don't contain invalid characters, if perf is affected in some cases.

You seem to be suggesting that we, in addition to RTE_TEL_STRING_VAL add a new telemetry value type for "strings not requiring any encoding by any protocols, JSON, HTML, SNMP, SOAP, InfluxDB, or whatever the future might bring", for performance reasons?

This would require that the telemetry providers call either rte_tel_data_plain_string() instead of rte_tel_data_string(), depending on the contents of the string, and damn well make sure they don't call rte_tel_data_plain_string() with strings containing special characters.

Alternatively, the rte_tel_data_string() function would need to scan for special characters and mark the string type accordingly.


I don't think it is a good idea modifying the lower layer of the telemetry library to support this.

I prefer to keep the choice of encoding or simple copying (assuming that simply copying provides a performance benefit) in the protocol specific encoding functions, such as rte_tel_json_str().

The performance cost of making the decision (to encode or not) in the protocol layer, i.e. in rte_tel_json_str(), is not higher than it is in the lower layer, i.e. in rte_tel_data_string(). And some protocols, e.g. SNMP, do allow a lot of special characters in strings without encoding.

> That
> said, in the testing I have done in the past:
> * for jobs like getting the nic stats or nic xstats, the vast majority
> of
>   the cycles spent in the telemetry function is doing PCI reads for the
>   stats
> * for simpler jobs e.g. getting list of ethdevs, the vast majority of
> the
>   time is actually being spent in the kernel socket handling
> 
> Based on what I've seen (and unfortunately I didn't note down the exact
> figures for above), its likely that even if performance of the
> interface is
> a concern we can probably add in some basic string encoding without
> really
> affecting things much.

Agree. Especially if we keep the implementation simple. E.g. we can reuse size limit from the unencoded for the encoded string.

> 
> /Bruce