def run_lat_cmd_sweep(nfp, outdir):
"""Run Latency tests to determine any cache or IO-MMU effects"""
twr = TableWriter(nfp.lat_fmt)
win_szs = [x * 1024 for x in [1, 4, 16, 256, 512]] + \
[x * 1024 * 1024 for x in [1, 1.5, 2, 3, 4, 8, 16, 32, 64]]
trans_szs = [8, 64]
access = nfp.FLAGS_RANDOM
for test_no in [nfp.LAT_CMD_RD, nfp.LAT_CMD_WRRD]:
for trans_sz in trans_szs:
out_name = "lat_cmd_sweep_"
out_name += "rd_" if test_no == nfp.LAT_CMD_RD else "wrrd_"
out_name += "seq_" if access == 0 else "rnd_"
out_name += "%02d" % trans_sz
twr.open(outdir + out_name, TableWriter.ALL)
twr.msg("\n\nPCIe CMD %s latency over different windows "
"sizes with %s access" %
("Read" if test_no == nfp.LAT_CMD_RD else "Write/Read",
"sequential" if access == 0 else "random"))
for flags in [
0, nfp.FLAGS_THRASH, nfp.FLAGS_WARM, nfp.FLAGS_HOSTWARM
]:
twr.sec()
for win_sz in win_szs:
_ = nfp.lat_test(twr, test_no, access | flags, win_sz,
trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_bw_dma_off(nfp, outdir):
"""Run Bandwidth tests to with different host offset"""
twr = TableWriter(nfp.bw_fmt)
win_sz = 8192
trans_szs = [64, 128, 256, 407, 416, 1024, 2048]
offsets = [0, 1, 2, 3, 4, 6, 8, 12, 16, 20, 25, 32, 41, 48, 63]
for flags in [0, nfp.FLAGS_HOSTWARM]:
for test_no in [nfp.BW_DMA_RD, nfp.BW_DMA_WR]:
out_name = "bw_dma_off_%s" % \
("rd" if test_no == nfp.BW_DMA_RD else "wr")
# With DDIO, no point in running cold/warm WRRD tests
if flags == 0 and test_no == nfp.LAT_DMA_WRRD:
continue
if flags == 0:
out_name += "_cold"
twr.open(outdir + out_name, TableWriter.ALL)
twr.msg("\nPCIe DMA %s Bandwidth with different host offset" %
("Read" if test_no == nfp.LAT_DMA_RD else "Write"))
for trans_sz in trans_szs:
twr.sec()
for off in offsets:
_ = nfp.bw_test(twr, test_no, flags, win_sz, trans_sz, off,
0)
twr.sec()
for off in offsets:
_ = nfp.bw_test(twr, test_no, flags, win_sz, trans_sz, 0,
off)
twr.close(TableWriter.ALL)
def run_lat_dma_off(nfp, outdir):
"""Run Latency tests to with different host offset"""
twr = TableWriter(nfp.lat_fmt)
flags = nfp.FLAGS_HOSTWARM
win_sz = 8192
trans_szs = [64, 128, 256, 407, 416, 1024, 2048]
offsets = [0, 1, 2, 3, 4, 6, 8, 12, 16, 20, 25, 32, 41, 48, 63]
for test_no in [nfp.LAT_DMA_RD, nfp.LAT_DMA_WRRD]:
out_name = "lat_dma_off_%s" % \
("rd" if test_no == nfp.LAT_DMA_RD else "wrrd")
twr.open(outdir + out_name, TableWriter.ALL)
twr.msg("\nPCIe DMA %s latency with different host offset" %
("Read" if test_no == nfp.LAT_DMA_RD else "Write/Read"))
for trans_sz in trans_szs:
twr.sec()
for off in offsets:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, off, 0)
twr.sec()
for off in offsets:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, 0, off)
twr.close(TableWriter.ALL)
def run_bw_dma_win_sweep(nfp, outdir):
"""Run Bandwidth tests with differnt windows sizes"""
twr = TableWriter(nfp.bw_fmt)
win_szs = [x * 1024 for x in [4, 16, 256, 512]] + \
[x * 1024 * 1024 for x in [1, 1.5, 2, 3, 4, 8, 16, 32, 64]]
trans_szs = [64, 128, 256, 512]
access = nfp.FLAGS_RANDOM
for test_no in [nfp.BW_DMA_RD, nfp.BW_DMA_WR]:
for trans_sz in trans_szs:
out_name = "bw_dma_win_sweep_"
out_name += "rd" if test_no == nfp.BW_DMA_RD else "wr"
out_name += "_rnd_"
out_name += "%02d" % trans_sz
twr.open(outdir + out_name, TableWriter.ALL)
twr.msg("\n\nPCIe DMA %s bandwidth over different windows "
"sizes with Random access" %
("Read" if test_no == nfp.BW_DMA_RD else "Write"))
for flags in [
0, nfp.FLAGS_THRASH, nfp.FLAGS_WARM, nfp.FLAGS_HOSTWARM
]:
twr.sec()
for win_sz in win_szs:
nfp.bw_test(twr, test_no, flags | access, win_sz, trans_sz,
0, 0)
twr.close(TableWriter.ALL)
def run_dbg_mem(nfp, outdir):
"""Debug memory, trying to hit the same cachelines over and over"""
twr = TableWriter(nfp.bw_fmt)
twr.open(outdir + "dbg_mem", TableWriter.ALL)
flags = nfp.FLAGS_HOSTWARM
test_no = nfp.BW_DMA_RD
trans_szs = [64, 128, 256, 512, 1024]
for trans_sz in trans_szs:
nfp.bw_test(twr, test_no, flags, trans_sz, trans_sz, 0, 0)
nfp.bw_test(twr, test_no, flags, 8192, trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_lat_cmd(nfp, outdir):
"""Run basic Latency tests for different sizes using the PCIe commands"""
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "lat_cmd_sizes", TableWriter.ALL)
twr.msg("\nPCIe CMD Latency with different transfer sizes")
win_sz = 4096
trans_szs = [4, 8, 16, 24, 32, 48, 64]
flags = nfp.FLAGS_HOSTWARM
for test_no in [nfp.LAT_CMD_RD, nfp.LAT_CMD_WRRD]:
twr.sec()
for trans_sz in trans_szs:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_lat_dma_byte(nfp, outdir):
"""Run basic Latency tests for different sizes using the PCIe commands"""
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "lat_dma_sizes_byte_inc", TableWriter.ALL)
twr.msg("\nPCIe DMA Latency with different transfer sizes")
win_sz = 8192
flags = nfp.FLAGS_HOSTWARM
for mid_sz in [256, 1024]:
trans_szs = range(mid_sz - 16, mid_sz + 16)
for test_no in [nfp.LAT_DMA_RD, nfp.LAT_DMA_WRRD]:
twr.sec()
for trans_sz in trans_szs:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_lat_dma(nfp, outdir):
"""Run basic Latency tests for different sizes using the PCIe commands"""
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "lat_dma_sizes", TableWriter.ALL)
twr.msg("\nPCIe DMA Latency with different transfer sizes")
win_sz = 8192
flags = nfp.FLAGS_HOSTWARM
trans_szs = [
4, 8, 16, 24, 32, 48, 64, 128, 256, 512, 768, 1024, 1280, 1520, 2048
]
for test_no in [nfp.LAT_DMA_RD, nfp.LAT_DMA_WRRD]:
twr.sec()
for trans_sz in trans_szs:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_dbg_lat(nfp, dma, write_read, win_sz, trans_sz, h_off, d_off, rnd,
long_run, cache_flags, outdir):
"""Run latency debug test"""
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "dbg_lat", TableWriter.ALL)
cdfwr = TableWriter(LAT_TEST_CDF_FMT, stdout=False)
cdfwr.open(outdir + "dbg_lat_details_cdf", TableWriter.ALL)
flags = cache_flags
if rnd:
flags |= nfp.FLAGS_RANDOM
if long_run:
flags |= nfp.FLAGS_LONG
if dma:
if write_read:
test_no = nfp.LAT_DMA_WRRD
else:
test_no = nfp.LAT_DMA_RD
else:
if write_read:
test_no = nfp.LAT_CMD_WRRD
else:
test_no = nfp.LAT_CMD_RD
lat_stats = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz, h_off,
d_off)
h_cyc = lat_stats.histo()
cdf_cyc = histo2cdf(h_cyc)
vals = sorted(cdf_cyc.keys())
cdfwr.sec("test=%s trans_sz=%d win_sz=%d" %
(nfp.TEST_NAMES[test_no], trans_sz, win_sz))
cdfwr.out((vals[0], nfp.cyc2ns(vals[0]), 0))
for val in vals:
cdfwr.out((val, nfp.cyc2ns(val), cdf_cyc[val]))
cdfwr.close(TableWriter.ALL)
twr.close(TableWriter.ALL)
def run_lat_cmd_off(nfp, outdir):
"""Run Latency tests to with different host offset"""
twr = TableWriter(nfp.lat_fmt)
win_szs = [4096, 8 * 1024 * 1024]
trans_szs = [8, 64]
for test_no in [nfp.LAT_CMD_RD, nfp.LAT_CMD_WRRD]:
out_name = "lat_cmd_off_%s" % \
("rd" if test_no == nfp.LAT_CMD_RD else "wrrd")
twr.open(outdir + out_name, TableWriter.ALL)
twr.msg("\nPCIe CMD %s latency with different host offset" %
("Read" if test_no == nfp.LAT_CMD_RD else "Write/Read"))
for trans_sz in trans_szs:
for win_sz in win_szs:
for flags in [0, nfp.FLAGS_HOSTWARM]:
twr.sec()
for off in [0, 1, 2, 3, 4, 6, 8, 16, 32, 48]:
_ = nfp.lat_test(twr, test_no, flags, win_sz, trans_sz,
off, 0)
twr.close(TableWriter.ALL)
def run_bw_dma_sz_sweep(nfp, outdir):
"""Run Bandwidth tests across different DMA sizes"""
twr = TableWriter(nfp.bw_fmt)
win_sz = 8192
# kinda assume MPS of 256
trans_szs = [
16, 32, 63, 64, 65, 127, 128, 129, 192, 255, 256, 257, 320, 384, 511,
512, 513, 576, 640, 704, 767, 768, 769, 832, 896, 960, 1023, 1024,
1025, 1279, 1280, 1281, 1535, 1536, 1537, 1791, 1792, 1793, 2047, 2048
]
flags = nfp.FLAGS_RANDOM | nfp.FLAGS_HOSTWARM
out_name = "bw_dma_sz_sweep"
twr.open(outdir + out_name, TableWriter.ALL)
for test_no in [nfp.BW_DMA_RD, nfp.BW_DMA_WR, nfp.BW_DMA_RW]:
twr.sec()
for trans_sz in trans_szs:
nfp.bw_test(twr, test_no, flags, win_sz, trans_sz, 0, 0)
twr.close(TableWriter.ALL)
def run_dbg_bw(nfp, wr_flag, rw_flag, win_sz, trans_sz, h_off, d_off, rnd,
cache_flags, outdir):
"""Run bandwidth debug test"""
if wr_flag and rw_flag:
raise Exception("Illegal combination of flags")
if wr_flag:
test_no = nfp.BW_DMA_WR
elif rw_flag:
test_no = nfp.BW_DMA_RW
else:
test_no = nfp.BW_DMA_RD
twr = TableWriter(nfp.bw_fmt)
twr.open(outdir + "dbg_bw", TableWriter.ALL)
flags = cache_flags
if rnd:
flags |= nfp.FLAGS_RANDOM
nfp.bw_test(twr, test_no, flags, win_sz, trans_sz, h_off, d_off)
twr.close(TableWriter.ALL)
def run_lat_details(nfp, outdir):
"""Run a longer test and perform some analysis"""
win_szs = [8192, 64 * 1024 * 1024]
common_flags = nfp.FLAGS_LONG | nfp.FLAGS_RANDOM
# Commands latencies
trans_szs = [8]
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "lat_cmd_details", TableWriter.ALL)
cdfwr = TableWriter(LAT_TEST_CDF_FMT, stdout=False)
cdfwr.open(outdir + "lat_cmd_details_cdf", TableWriter.ALL)
raw = open(outdir + "lat_cmd_details_raw.dat", 'w')
twr.msg("\nPCIe CMD latencies with more details")
for test_no in [nfp.LAT_CMD_RD, nfp.LAT_CMD_WRRD]:
twr.sec()
for trans_sz in trans_szs:
for win_sz in win_szs:
for flags in [common_flags, common_flags | nfp.FLAGS_HOSTWARM]:
lat_stats = nfp.lat_test(twr, test_no, flags, win_sz,
trans_sz, 0, 0)
h_cyc = lat_stats.histo()
cdf_cyc = histo2cdf(h_cyc)
vals = sorted(cdf_cyc.keys())
cdfwr.sec("test=%s trans_sz=%d win_sz=%d cache=%s" %
(nfp.TEST_NAMES[test_no], trans_sz, win_sz,
"hwarm" if flags & nfp.FLAGS_HOSTWARM \
else "cold"))
cdfwr.out((vals[0], nfp.cyc2ns(vals[0]), 0))
for val in vals:
cdfwr.out((val, nfp.cyc2ns(val), cdf_cyc[val]))
# write raw data
raw.write(
"# %s %s Winsz=%d trans_sz=%d (values in ns)\n" %
(nfp.TEST_NAMES[test_no], "Warm" if flags
& nfp.FLAGS_HOSTWARM else "Cold", win_sz, trans_sz))
for t in lat_stats.list:
raw.write("%.0f\n" % nfp.cyc2ns(t))
raw.write("\n\n")
cdfwr.close(TableWriter.ALL)
twr.close(TableWriter.ALL)
raw.close()
# DMA latencies
trans_szs = [64, 2048]
twr = TableWriter(nfp.lat_fmt)
twr.open(outdir + "lat_dma_details", TableWriter.ALL)
cdfwr = TableWriter(LAT_TEST_CDF_FMT, stdout=False)
cdfwr.open(outdir + "lat_dma_details_cdf", TableWriter.ALL)
raw = open(outdir + "lat_dma_details_raw.dat", 'w')
twr.msg("\nPCIe DMA latencies with more details")
for test_no in [nfp.LAT_DMA_RD, nfp.LAT_DMA_WRRD]:
twr.sec()
for trans_sz in trans_szs:
for win_sz in win_szs:
for flags in [common_flags, common_flags | nfp.FLAGS_HOSTWARM]:
lat_stats = nfp.lat_test(twr, test_no, flags, win_sz,
trans_sz, 0, 0)
h_cyc = lat_stats.histo()
cdf_cyc = histo2cdf(h_cyc)
vals = sorted(cdf_cyc.keys())
cdfwr.sec("test=%s trans_sz=%d win_sz=%d cache=%s" %
(nfp.TEST_NAMES[test_no], trans_sz, win_sz,
"hwarm" if flags & nfp.FLAGS_HOSTWARM \
else "cold"))
cdfwr.out((vals[0], nfp.cyc2ns(vals[0]), 0))
for val in vals:
cdfwr.out((val, nfp.cyc2ns(val), cdf_cyc[val]))
# write raw data
raw.write(
"# %s %s Winsz=%d trans_sz=%d (values in ns)\n" %
(nfp.TEST_NAMES[test_no], "Warm" if flags
& nfp.FLAGS_HOSTWARM else "Cold", win_sz, trans_sz))
for t in lat_stats.list:
raw.write("%.0f\n" % nfp.cyc2ns(t))
raw.write("\n\n")
cdfwr.close(TableWriter.ALL)
twr.close(TableWriter.ALL)
raw.close()