The node exporter runs unprivileged, so it cannot statfs any filesystems
under this directory causing log spam. In addition there tends to be
high churn in the filesystems here (as it's basically application
monitoring) which can cause high cardinaltiy and in one case caused
Prometheus's index symbol table to get very large.
Accordingly this should be ignored to reduce log spam and avoid
performance issues. The filesystems themselves can in principle be
monitored via container oriented exporters, and the underlying
filesystems will still be monitored.
* Unify CPU collector conventions
Add a common CPU metric description.
* All collectors use the same `nodeCpuSecondsDesc`.
* All collectors drop the `cpu` prefix for `cpu` label values.
* Fix subsystem string in cpu_freebsd.
* Fix Linux CPU freq label names.
* Improve stat linux metric names.
cpu is no longer used.
* node_cpu -> node_cpu_seconds_total for Linux
* Improve filesystem metric names with units
* Improve units and names of linux disk stats
Remove sector metrics, the bytes metrics cover those already.
* Infiniband counters should end in _total
* Improve timex metric names, convert to more normal units.
See
3c073991eb/kernel/time/ntp.c (L909)
for what stabil means, looks like a moving average of some form.
* Update test fixture
* For meminfo metrics that had "kB" units, add _bytes
* Interrupts counter should have _total
* Move FreeBSD/DragonflyBSD out of meminfo add kvm.
This gives us SwapUsed, and everything under one roof.
* Fix typos per review.
* Update to use newer API.
* Remove premature optimization per PR feedback.
* Implements meminfo collector for OpenBSD
This is a rework of #151.
* Fix CGO import
* Add some useful metrics
* Rename total -> size for normalization
* remove injection hook for textfile metrics, convert them to prometheus format
* add support for summaries
* add support for histograms
* add logic for handling inconsistent labels within a metric family for counter, gauge, untyped
* change logic for parsing the metrics textfile
* fix logic to adding missing labels
* Export time and error metrics for textfiles
* Add tests for new textfile collector, fix found bugs
* refactor Update() to split into smaller functions
* remove parseTextFiles(), fix import issue
* add mtime metric directly to channel, fix handling of mtime during testing
* rename variables related to labels
* refactor: add default case, remove if guard for metrics, remove extra loop and slice
* refactor: remove extra loop iterating over metric families
* test: add test case for different metric type, fix found bug
* test: add test for metrics with inconsistent labels
* test: add test for histogram
* test: add test for histogram with extra dimension
* test: add test for summary
* test: add test for summary with extra dimension
* remove unnecessary creation of protobuf
* nit: remove extra blank line
Linux "guest" metrics for VMs are already accounted for in node_cpu
`user` and `nice` metrics. Separate these into their own metric to
avoid duplication of data.
* cpu: Support processor-less (memory-only) NUMA nodes
Processor-less (memory-only) NUMA nodes exist e.g. in systems that use
Intel Optane drives for RAM expansion using Intel Memory Drive
Technology (IMDT).
IMDT RAM expansion supports two modes:
* "Unify Remote Memory domains": present a processor-less (memory-only)
NUMA domain, which is the default
* "Expand local memory domains": to expand each processor’s memory domain
with a portion of the memory made available by Optane and IMDT
This commit fixes a crash in the first case (when "cpulist" is empty).
Here's an example of such a system:
$ numastat -m|head -n5
Per-node system memory usage (in MBs):
Node 0 Node 1 Node 2 Total
--------------- --------------- --------------- ---------------
MemTotal 118239.56 130816.00 464384.00 713439.56
$ for i in {0..2}; do echo -n "$i: " ; cat /sys/bus/node/devices/node$i/cpulist ; done
0: 0-7,16-23
1: 8-15,24-31
2:
$ /opt/vsmp/bin/vsmpversion -vvv
Memory Drive Technology: 8.2.1455.74 (Sep 28 2017 13:09:59)
System configuration:
Boards: 3
1 x Proc. + I/O + Memory
2 x NVM devices (Intel SSDPED1K375GAQ)
Processors: 2, Cores: 16, Threads: 32
Intel(R) Xeon(R) CPU E5-2667 v4 @ 3.20GHz Stepping 01
Memory (MB): 713472 (of 977450), Cache: 251416, Private: 12562
1 x 249088MB [262036/ 678/12270]
1 x 232192MB [357707/125369/ 146] 82:00.0#1
1 x 232192MB [357707/125369/ 146] 83:00.0#1
* cpu: rename some variables (pkg => node)
* cpu: Use %v not %q in log.Debugf() format strings
* Update golang.org/x/sys/unix
This allows to use simplified string conversion of Utsname members.
* Simplify Utsname string conversion
Use Utsname from golang.org/x/sys/unix which contains byte array
instead of int8/uint8 array members. This allows to simplify the string
conversions of these members.
* Correct buffer_bytes > INT_MAX on BSD/amd64.
The sysctl vfs.bufspace returns either an int or a long, depending on
the value. Large values of vfs.bufspace will result in error messages
like:
couldn't get meminfo: cannot allocate memory
This will detect the returned data type, and cast appropriately.
* Added explicit length checks per feedback.
* Flatten Value() to make it easier to read.
* Simplify per feedback.
* Fix style.
* Doc updates.
The github.com/beevik/ntp package was recently updated with some
API changes that broke node_exporter. This commit fetches the
latest version of the ntp package and brings node_exporter in
line with the latest API.
* Move NodeCollector into package collector
* Refactor collector enabling
* Update README with new collector enabled flags
* Fix out-of-date inline flag reference syntax
* Use new flags in end-to-end tests
* Add flag to disable all default collectors
* Track if a flag has been set explicitly
* Add --collectors.disable-defaults to README
* Revert disable-defaults flag
* Shorten flags
* Fixup timex collector registration
* Fix end-to-end tests
* Change procfs and sysfs path flags
* Fix review comments
This collector is based on adjtimex(2) system call. The collector returns
three values, status if time is synchronised, offset to remote reference,
and local clock frequency adjustment.
Values are taken from kernel time keeping data structures to avoid getting
involved how the synchronisation is implemented. By that I mean one should
not care if time is update using ntpd, systemd.timesyncd, ptpd, and so on.
Since all time sync implementation will always end up telling to kernel what
is the status with time one can simply omit the software in between, and
look results of the syncing. As a positive side effect this makes collector
very quick and conceptually specific, this does not monitor availability of
NTP server, or network in between, or dns resolution, and other unrelated
but necessary things.
Minimum set of values to keep eye on are the following three:
The node_timex_sync_status tells if local clock is in sync with a remote
clock. Value is set to zero when synchronisation to a reliable server
is lost, or a time sync software is misconfigured.
The node_timex_offset_seconds tells how much local clock is off when
compared to reference. In case of multiple time references this value
is outcome of RFC 5905 adjustment algorithm. Ideally offset should be
close to zero, and it depends about use case how large value is
acceptable. For example a typical web server is probably fine if offset
is about 0.1 or less, but that would not be good enough for mobile phone
base station operator.
The node_timex_freq tells amount of adjustment to local clock tick
frequency. For example if offset is one second and growing the local
clock will need instruction to tick quicker. Number value itself is not
very important, and occasional small adjustments are fine. When
frequency is unusually in stable one can assume quality of time stamps
will not be accurate to very far in sub second range. Obviously
explaining why local clock frequency behaves like a passenger in roller
coaster is different matter. Explanations can vary from system load, to
environmental issues such as a machine being physically too hot.
Rest of the measurements can help when debugging. If you run a clock server
do probably want to collect and keep track of everything.
Pull-request: https://github.com/prometheus/node_exporter/pull/664
* Add metrics from SNTPv4 packet to ntp collector & add ntpd sanity check
1. Checking local clock against remote NTP daemon is bad idea, local
ntpd acting as a client should do it better and avoid excessive load on
remote NTP server so the collector is refactored to query local NTP
server.
2. Checking local clock against remote one does not check local ntpd
itself. Local ntpd may be down or out of sync due to network issues, but
clock will be OK.
3. Checking NTP server using sanity of it's response is tricky and
depends on ntpd implementation, that's why common `node_ntp_sanity`
variable is exported.
* `govendor add golang.org/x/net/ipv4`, it is dependency of github.com/beevik/ntp
* Update github.com/beevik/ntp to include boring SNTP fix
* Use variable name from RFC5905
* ntp: move code to make export of raw metrics more explicit
* Move NTP math to `github.com/beevik/ntp`
* Make `golint` happy
* Add some brief docs explaining `ntp` #655 and `timex` #664 modules
* ntp: drop XXX comment that got its decision
* ntp: add `_seconds` suffix to relevant metrics
* Better `node_ntp_leap` comment
* s/node_ntp_reftime/node_ntp_reference_timestamp_seconds/ as requested by @discordianfish
* Extract subsystem name to const as suggested by @SuperQ
* cpu: Metric 'package_throttles_total' is per package.
'package_throttles_total' is per package, not per cpu. This also reduces
the total number of cpu time series a lot (esp for multi core cpus).
* cpu: Better handling of a cpulist edge-case.
* cpu: Extract the package number from the directory name.
Do not rely on the range index.
* cpu: Add package_throttle_count for node0 cpu1
This file must be ignored by the cpu collector.
This avoids issues with integer overflows on 32-bit architectures. The
Prometheus data format is float64, so regardless of the architecture we
should handle large numbers.
Fixes#629.
* Add bcache collector for Linux
This collector gathers metrics related to the Linux block cache
(bcache) from sysfs.
* Removed commented out code
* Use project comment style
* Add _sectors to metric name to indicate unit
* Really use project comment style
* Rename bcache.go to bcache_linux.go
* Keep collector namespace clean
Rename:
- metric -> bcacheMetric
- periodStatsToMetrics -> bcachePeriodStatsToMetric
* Shorten slice initialization
* Change label names to backing_device, cache_device
* Remove five minute metrics (keep only total)
* Include units in additional metric names
* Enable bcache collector by default
* Provide metrics in seconds, not nanoseconds
* remove metrics with label "all"
* Add fixtures, update end-to-end for bcache collector
* Move fixtures/sys into tar.gz
This changeset moves the collector/fixtures/sys directory into
collector/fixtures/sys.tar.gz and tweaks the Makefile to unpack the
tarball before tests are run.
The reason for this change is that Windows does not allow colons in a
path (colons are present in some of the bcache fixture files), nor can
it (out of the box) deal with pathnames longer than 260 characters
(which we would be increasingly likely to hit if we tried to replace
colons with longer codes that are guaranteed not the turn up in regular
file names).
* Add ttar: plain text archive, replacement for tar
This changeset adds ttar, a plain text replacement for tar, and uses it
for the sysfs fixture archive. The syntax is loosely based on tar(1).
Using a plain text archive makes it possible to review changes without
downloading and extracting the archive. Also, when working on the repo,
git diff and git log become useful again, allowing a committer to verify
and track changes over time.
The code is written in bash, because bash is available out of the box on
all major flavors of Linux and on macOS. The feature set used is
restricted to bash version 3.2 because that is what Apple is still
shipping.
The programm also works on Windows if bash is installed. Obviously, it
does not solve the Windows limitations (path length limited to 260
characters, no symbolic links) that prompted the move to an archive
format in the first place.
* Add diskstats collector for Darwin
* Update year in the header
* Update README.md
* Add github.com/lufia/iostat to vendored packages
* Change stats to follow naming guidelines
* Add a entry of github.com/lufia/iostat into vendor.json
* Remove /proc/diskstats from description
* Add qdisc collector for Linux
This collector gathers basic queueing discipline metrics via netlink,
similarly to what `tc -s qdisc show` does.
* qdisc collector: nl-specific code moved, names fixed
- netlink-specific parts moved to github.com/ema/qdisc
- avoid using shortened names
- counters renamed into XXX_total
* Get rid of parseMessage error checking leftover
* Add github.com/ema/qdisc to vendored packages
* Update help texts and comments
* Add qdisc collector to README file
* qdisc collector end-to-end testing
* Update qdisc dependency to latest version
Update github.com/ema/qdisc dependency to revision 2c7e72d, which
includes unit testing.
* qdisc collector: rename "iface" label into "device"
According to Mellanox, it is standard practice that the port_xmit_data and port_rcv_data
files are split into 4 lanes. To get the actual transmit and receive values for each
port, the metric needs to be multiplied by 4.
Signed-Off-By: Robert Clark <robert.d.clark@hpe.com>
* silently ignore nonexisting bonding_masters file
Add an empty fixtures dir without a bonding_masters file to test.
* Moved the check to the Update() method
Dropped the empty test dir.
Since Go 1.8 32bit MIPS Big/Little Endian are supported assuming the
target runs Linux and the kernel either emulates an FPU or can access
the CPU one.
This allows the node_collector to build for mips and mipsle opening up
the possibility of running it on things like home routers
(DD-|Open|ASUS-)Wrt firmware usually has the necessary bits in place.
* Implement commonalities and linux support for ARP collection
* Add ARP collector to fixtures and run as part of e2e tests
* Bubble up scanner errors
* Use single return values where it makes sense
* Add missing annotation
* Move arp_common into arp_linux
* Add license header to arp_linux.go
* Address initial feedback
* Use strings.Fields instead of strings.Split
* Deal with scanner.Err() rather than throwing away errors
* Check for scan errors in-line before interacting with the entries map
* Don't interact with potentially empty text from scan
* Check for scan errors outside the scan loop
* Add comment about moving procfs parsing
* Add more direct comment
* Update initialism style to match go style guide
* Put function args on the same line
* Add TODO in front of comment about procfs extraction
* Guard against strings.Fields returning an empty slice
* Be more defensive about ARP table format and use upcase more broadly
* Enable the ARP collector by default
* Add ARP collector to the README
* Remove 'entry'
Instead of maintaining a counter metric for device errors in memory,
this change exports a gauge and uses const metrics to avoid leaking
metrics for unmounted filesystems.
Older versions of the OFED drivers contain 64-bit variants of the port counters and are located in a directory named 'counters_ext'. This patch includes these older metrics that have since been deprecated with OFED 4.0.
Signed-Off-By: Robert Clark <robert.d.clark@hpe.com>
In case a metric file within the InfiniBand collector doesn't exist, skip the metric in order to allow collection of the remaining valid InfiniBand metrics.
Signed-Off-By: Robert Clark <robert.d.clark@hpe.com>
Matt Layher