The latency tracking is done with ranges of bucket arrays.
The bucket for any given I/O is determined solely by TSC
deltas - any translation to microseconds is only done after
the test is finished and statistics are printed.
Each range has a number of buckets determined by a
NUM_BUCKETS_PER_RANGE value which is currently set to 128.
The buckets in ranges 0 and 1 each map to one specific TSC
delta. The buckets in subsequent ranges each map to twice
as many TSC deltas as buckets in the previous range:
Range 0: 1 TSC each - 128 buckets cover deltas 0 to 127
Range 1: 1 TSC each - 128 buckets cover deltas 128 to 255
Range 2: 2 TSC each - 128 buckets cover deltas 256 to 511
Range 3: 4 TSC each - 128 buckets cover deltas 512 to 1023
Range 4: 8 TSC each - 128 buckets cover deltas 1024 to 2047
Range 5: 16 TSC each - 128 buckets cover deltas 2048 to 4095
etc.
While here, change some variable names and usage
messages to differentiate between the existing latency
tracking via vendor-specific NVMe log pages on Intel
NVMe SSDs, and the newly added latency tracking done
in software.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I299f1c1f6dbfa7ea0e73085f7a685e71fc687a2b
These variables are not needed outside of this file.
Change-Id: I577db5140fb4156100e72e6219c8121304a9b137
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This patch is causing intermittent errors in the automated tests, so
reverting for now until this can be resolved.
This reverts commit 1d9891989c.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I237e31b1b8a1a1e28262851ae0a21cd7345f0f1b
Fix up all existing spacing errors in comments and add an automated
check for patterns like /*comment*/.
Change-Id: I28f61c93612dc0f8aed66bd509da78e91ea9737e
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The reaon is that rte_eal_init called in spdk_env_init binds
the master thread on master core, we need to restore the
affinity.
Change-Id: I4eb2d36b74404c1af6d242c4fd5c403bd3d5036c
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
This patch will be prepare to support multi threads.
Change-Id: I2234ae1d564c9cf17b92c85fd0a414adaa639bf2
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Also adds the automation test case for using remote
NVMe devices exported by NVMe-oF target.
Change-Id: I2b839a4eeec33d5b0c30d654e6013ad8c7949e23
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
This is required by libaio. Previously, buffers were aligned to 512
bytes, but 4K devices need 4K-aligned buffers.
Change-Id: I96080e72dc77e0e72f426f7c9fe98b6724f66e1b
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Several examples have a function to associate workers
with threads. Simplify that algorithm.
This seems to just shift some of the complexity
from register_workers down to main, but in the long
run the DPDK threading will get abstracted into
env as well and greatly simplify that part.
Change-Id: Ic106dde58fa5351a1ce0a058161b08062e121d3b
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The SPDK fio plugin only supports jobs as threads, not separate
processes. So log an error and bail if --thread=1 is not
specified.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I318f924e0b7fbb76f11427a447dd7c5fc0ebc221
fgets guarantees that strings returned are NULL terminated,
so delete logic that attempts to insert a NULL terminator.
Change-Id: Ibeadcda9af431091b04daff41596d5738419b43f
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
The previous implementation tried to maintain a ring buffer of completed
I/O retrieved from ->getevents() that had not yet been retrieved by
->event(), but this is not necessary. None of the other FIO engines
implement this sort of ring buffer, and the only code that calls
->getevents() will always retrieve the full number of events via
->event() afterwards.
Instead, just count how many I/Os have been completed within each
->getevents() call in iocq_count, and reset iocq_count on each new
->getevents().
Change-Id: I2776c3a5c89a0d2b24478328bc138ff7a9fb5bdd
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
SPDK_COUNTOF works like sizeof, except it returns the number of elements
in an array instead of the number of bytes.
Change-Id: I38ff4dd3485ed9b630cc5660ff84851d0031911f
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Since we know the queue depth that we will be using during the test,
request that as the queue size when attaching to NVMe controllers.
Reserve one extra queue entry above the expected queue depth since NVMe
queues must always have one entry free to distinguish between queue
empty and queue full cases.
Change-Id: I809982207edb4894148aec09b10c4e2de4a040d3
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
By default, all SPDK applications will not share memory.
To share memory, start the applications with the same
shared memory id.
Change-Id: Ib6180369ef0ed12d05983a21d7943e467402b21a
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
The "idle" status has 2 meanings for IOAT kernel performace
test module dmaperf: performance test task has finished or
no performance task at all. For the lastest kernel(4.8), the
userland tool will get the "idle" status before the kernel
changed to "running", but for some older kernels(3.17,4.0)
the test tool run correctly. So add a sleep syscall before
to get the status will fix it.
Change-Id: Ia236416607c77b3b8689933fe71ce53a783a04cc
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Map the DPDK hugepage virtual address space to an area that should not
interfere with randomized mmap() addresses.
Change-Id: Iffc657858f861fc1316f77b68f9f121167d604b1
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
The previous performance caculation is wrong,
which is smaller than the per channel performance,
so fix it with the average performance data.
Change-Id: I40580e5f70fb4273fa080bbdfb17ce85436446aa
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
If no channels are available, don't try to run the test.
Change-Id: I4c40635a3da598064da7c94b3c7960a83ba25f8f
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Previously each core can only leverage one ioat channel.
With this patch, we can support the following features:
(1) Users can input the number of ioat channel to be tested.
(2) If the number of CPU cores is c, the total ioat channels
are n, each cpu core can use n/c channels.
Change-Id: I6b2179d3d633cb0999a3f4c5f40d4605c8cebc45
Signed-off-by: Ziye Yang <ziye.yang@intel.com>
The kernel module parameter (ioat_ring_alloc_order) was removed in
kernel commit cd60cd96137f6cb3ea82cace9225626619e7a52d (dmaengine:
IOATDMA: Removing descriptor ring reshape).
This parameter was only used to issue a warning about a too-large queue
depth, so just hard-code the current kernel's limit.
Change-Id: Ib3b32851f538164590b82cb347b02c1c3b9c0106
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Fix the getopt string for the -h switch so that it does not expect an
argument.
Change-Id: I8ea86018d6ee2c61ad698eead9718b2babfe6791
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
This allows the user to connect to multiple remote NVMe-oF targets or to
specify multiple specific PCIe device addresses to test.
Change-Id: I05b2072b8aa1480891b37b17b5207369344b617d
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
