From mboxrd@z Thu Jan 1 00:00:00 1970 Content-Type: multipart/mixed; boundary="===============0388155914384429556==" MIME-Version: 1.0 From: Walker, Benjamin Subject: Re: [SPDK] Tool to monitor disk activity Date: Wed, 12 Oct 2016 17:40:21 +0000 Message-ID: <1476294019.63592.86.camel@intel.com> In-Reply-To: CAG=8psVWYvBxCKp6ikZRr7g8oiS_PaEbD=HvfKrq9u_+=f39-Q@mail.gmail.com List-ID: To: spdk@lists.01.org --===============0388155914384429556== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Wed, 2016-10-12 at 04:02 +0000, Naoki Iwakami wrote: > Hello, > = > I am looking for a way to monitor disk activity while I access an SSD via > SPDK. > For example, any disk activity does not appear in vmstat or iostat > while I am running SPDK perf tool. > I understand it is expected, but I wonder if there is any tool > available to replace iostat for SPDK. > Or is there any SPDK API method that helps probing disk activities? > = I think this is an important problem to solve as SPDK matures into a soluti= on used in production deployments, and we've been giving it some thought. Toda= y, our only available tool is our tracing library (include/spdk/trace.h, lib/trace), which is an efficient mechanism to add small trace points to the code. The trace points are placed in a shared memory region that can be map= ped by other tools to do analysis. We're currently investigating either expandi= ng this (adding well documented tracepoints instead of the ad hoc ones we have= in there now), or changing the format. At some point someone really needs to w= rite a simplified iostat-like tool that consumes these tracepoints and displays I/Ops and bandwidth. We have a very basic example of how this would work (app/iscsi_t= op) that displays live connection statistics for the iSCSI target, so maybe tha= t's a reasonable starting point. It would be awesome if someone from the community took on the task of formalizing the tracepoints and writing the tool! We're also working on multi-process support for the NVMe driver, which means multiple processes will be able to create I/O queues for the same NVMe device. This could potentially allow tools to attach to an active SPDK process and send comman= ds directly to the SSD to query performance metrics, if the SSD provides them. I think our long term vision= for this feature is to enable support for configuration and management of SPDK-= owned = NVMe SSDs with the standard nvme-cli tool, but we're quite a ways off yet. We're very open to ideas and feedback in this area if anyone has anything t= o say on the subject! > Thank you, --===============0388155914384429556==--