C741 runs SPS - ME runs on W790
SPS is involved but can't be modded. more easy in BIOS.
ES Xeon Discussion
- Thread starter britinpdx
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
SPS is involved but can't be modded. more easy in BIOS.
With disabled cores, does this slow down QYFP's cache access, or remove total cache? So say 56 disabled to 28 cores, still has access to full 105mb of cache? And is there a bios option to stop accessing cache from another socket? Or is the cache only accessed on the utilized socket cores?
why do you not want the data from the much faster L3 ?
if you change the data in the Real RAM, it must be declared invalid if this location is stored in some L3 cache,
otherwise all other cores using the L3 cache work with wrong(old) data.
I wonder which bios are you using. I can't find any F0* version MS33-CPA BIOS anywhere.
My solution was to hybridise the AR0 and CPA bios to get support for DCPMM, at the cost of two MCIO 8i's not working.
May I ask what your PMEM model string is? For example, something like NMC2XXD256GPS.
please note: it's generally advised to put larger images under spoiler tag (although initially I was guilty to cause these troubles too). thanks.
ES - possible only memory mode supported(at least older gens have this limitation)
Gigabyte MS33-AR0 + EMR Q2SR + 8 x DDR5 64Gb, 4800 MHz (Kingston Server Premier KSM48R40BD4TMM-64HMR).
BIOS R07 (tweak RolloZ170): Standart Perfomance, Virtual NUMA = Disable (2 nodes).
AlmaLinux 9.5, ClickHouse server version 25.1.1.3390 (official build).
echo "SELECT * FROM system.numbers LIMIT 100_000_000 OFFSET 100_000_000 SETTINGS max_threads = 128, max_insert_threads = 128" | ./clickhouse-benchmark --host=localhost --port=9000 -i 10
Number of processed queries.
QPS: How many queries the server performed per second
RPS: How many rows the server reads per second
MiB/s: How many mebibytes the server reads per second
result RPS: How many rows placed by the server to the result of a query per second
result MiB/s. How many mebibytes placed by the server to the result of a query per second
Percentiles of queries execution time.
BIOS R07 (tweak RolloZ170): Standart Perfomance, Virtual NUMA = Disable (2 nodes).
AlmaLinux 9.5, ClickHouse server version 25.1.1.3390 (official build).
echo "SELECT * FROM system.numbers LIMIT 100_000_000 OFFSET 100_000_000 SETTINGS max_threads = 128, max_insert_threads = 128" | ./clickhouse-benchmark --host=localhost --port=9000 -i 10
Number of processed queries.
QPS: How many queries the server performed per second
RPS: How many rows the server reads per second
MiB/s: How many mebibytes the server reads per second
result RPS: How many rows placed by the server to the result of a query per second
result MiB/s. How many mebibytes placed by the server to the result of a query per second
Percentiles of queries execution time.
Impressive, but at the same time 100mil+ rows per second might not be something break-through/"amazing".
I know it may not be directly subject for ES but while it's hot:
1) Do you have any bench at least from some reference highly-granular single-table (no joins) of real DB (preferably fitting inMemory {so it's probably 400GB+}, desirably numerical-only columns something like key-key-[bigint]value, no strings) with billions of rows to check how it works with ColumnStore index (to avoid impact by some storage bottlenecks)? - just asking "in case", something like this can be actually "too much" anyway.
2) Does Q2SR has IAA accelerators (how many units CPU has?) and ClikHouse configured to use it (e.g. via QPL) and AVX-512 ?
I'm not familiar specifically with CH (mostly MS-focused) but AFAIK some DBs on a back-engine side can "optimize" execution plan if there's no need to fetch results back to client leading to crazy high scores (although your note says "placed by the server to the result" suggesting benchmark takes care of this).
3) What's the "effective clock speed" on CPU during benchmark?
p.s. pretty sure mainstream DBs aren't programmed yet to take advantage of AMX?
Thanks.
I know it may not be directly subject for ES but while it's hot:
1) Do you have any bench at least from some reference highly-granular single-table (no joins) of real DB (preferably fitting inMemory {so it's probably 400GB+}, desirably numerical-only columns something like key-key-[bigint]value, no strings) with billions of rows to check how it works with ColumnStore index (to avoid impact by some storage bottlenecks)? - just asking "in case", something like this can be actually "too much" anyway.
2) Does Q2SR has IAA accelerators (how many units CPU has?) and ClikHouse configured to use it (e.g. via QPL) and AVX-512 ?
I'm not familiar specifically with CH (mostly MS-focused) but AFAIK some DBs on a back-engine side can "optimize" execution plan if there's no need to fetch results back to client leading to crazy high scores (although your note says "placed by the server to the result" suggesting benchmark takes care of this).
3) What's the "effective clock speed" on CPU during benchmark?
p.s. pretty sure mainstream DBs aren't programmed yet to take advantage of AMX?
Thanks.
Last edited:
Q2SR has 4 IAX units.
PCI Address (Bus Device:Function) Number: 231:2:0
PCI Address (Bus Device:Function) Number: 236:2:0
PCI Address (Bus Device:Function) Number: 241:2:0
PCI Address (Bus Device:Function) Number: 246:2:0
Last edited:
What do you think, should something be done about the fact that MS73-HB according to the specification supports TDP up to 270W per processor? Q2SR TDP 350W
Thanks.
RolloZ170, is EMR-SP QS processor seems like "Genue Intel 0000" in tests, or like a production CPU?
QS is production CPU but ES bit set. you see full brand string, not "0000"
but some sellers state they sell QS, but deliver "0000" ES ones.
seller of QS show production units in the're listing, strange new worlds.