I must admit that I didn't get the time to go and check sensors project yet. But as for HAR, Peter it's exactly that: the aim is to give a higher abstraction for the hardware counters. As you said the individual counters are hard to understand and normaly to get a glimps of a notion like flops, or data cache misses, you would need multiple counters and do some mathematical operations to get a human exploitable result. That's exactly what HAR does. We have chosen a dozen of higher level metrics and for each different CPU, we implement what counters to read and what operations to do to get that metric. For example if you are running a HPC programme and would like to see the Mflops you don't really have to care if it's running on a SPARC IIIi, SPARC IV, Intel Xeon or etc. HAR takes care of the CPU specificities
Peter Tribble a ?crit : > On 1/14/08, Cynthia McGuire <cindi at sun.com> wrote: > >> This is exactly what the Sensor Abstraction project >> (http://opensolaris.org/os/project/sensors) is suppose to address. To >> get engaged with that project, subscribe and participate in the >> discussions at http://opensolaris.org/os/community/fm >> > > I'm not sure that I follow the connection. As I understand it > sensors is about abstracting fault telemetry; the HAR proposal > is about abstracting hardware-specific performance data. > Similar in that they are an abstraction layer above data sources > that are almost impossible to understand individually, but > with different underlying data and consumers. Or am I missing > something? > > I would love to see a higher level of abstraction for the hardware > performance counters. > > >> Amir Javanshir wrote: >> >>> Hi all >>> >>> I would like to submit a new project request and hopefully get a sponsor. >>> >>> Project Name: >>> HAR (Hardware Activity Reporter) >>> >>> Project Description >>> >>> The goal of this project is to develop a performance monitoring tool >>> that samples hardware counters in the system (primarily cpu counters, >>> but could extend to bus counters) to produce higher-level metrics (eg, >>> mips, flops, cache miss, stall rate, bus utilization) which can guide >>> a bottleneck analysis and performance tuning process. There exists a >>> first release of HAR, developed internally at Sun, that only looks at >>> cpu counters and supports currently only UltraSPARC1-4 (not 4+) & >>> Pentium3 systems. The open-sourcing of the existing HAR code has been >>> approved by Sun. There are a few users of HAR out there that keep >>> asking for newer releases (eg eBay) and support for recent hardware. >>> The purpose of open-sourcing HAR is to provide a proper place for the >>> continued development & dissemination of HAR. The primary task will be >>> to port HAR to Solaris 10 and libcpc2 so Niagara, AMD and modern Intel >>> processors can be supported. >>> >>> Related Projects >>> >>> There are no known related projects or dependencies. HAR will build >>> directly on top of libcpc2 for accessing cpu counters. Access to bus >>> counters has not been researched yet. >>> >>> Current Team >>> >>> We are already 3 engineers working on this subject >>> >>> Cheers, >>> Amir Javanshir >>> >>> >>> This message posted from opensolaris.org >>> _______________________________________________ >>> request-sponsor mailing list >>> request-sponsor at opensolaris.org >>> >> _______________________________________________ >> request-sponsor mailing list >> request-sponsor at opensolaris.org >> >> > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://mail.opensolaris.org/pipermail/observability-discuss/attachments/20080117/32c2d9f2/attachment.html>
