mistranslation: 0.766
permissions: 0.757
graphic: 0.736
user-level: 0.730
semantic: 0.729
virtual: 0.718
debug: 0.701
architecture: 0.696
performance: 0.679
assembly: 0.671
TCG: 0.670
register: 0.661
hypervisor: 0.649
PID: 0.649
peripherals: 0.639
device: 0.634
KVM: 0.623
risc-v: 0.611
arm: 0.605
boot: 0.600
vnc: 0.595
VMM: 0.582
network: 0.579
socket: 0.573
ppc: 0.568
kernel: 0.551
x86: 0.532
i386: 0.514
files: 0.475

Configured Memory access latency and bandwidth not taking effect

I was trying to configure latencies and bandwidths between nodes in a NUMA emulation using QEMU 5.0.0.

Host : Ubuntu 20.04 64 bit
Guest : Ubuntu 18.04 64 bit
 
The machine configured has 2 nodes. Each node has 2 CPUs and has been allocated 3GB of memory. The memory access latencies and bandwidths for a local access (i.e from initiator 0 to target 0, and from initiator 1 to target 1) are set as 40ns and 10GB/s respectively. The memory access latencies and bandwidths for a remote access (i.e from initiator 1 to target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s respectively. 

The command line launch is as follows.

sudo x86_64-softmmu/qemu-system-x86_64  \
-machine hmat=on \
-boot c \
-enable-kvm \
-m 6G,slots=2,maxmem=7G \
-object memory-backend-ram,size=3G,id=m0 \
-object memory-backend-ram,size=3G,id=m1 \
-numa node,nodeid=0,memdev=m0 \
-numa node,nodeid=1,memdev=m1 \
-smp 4,sockets=4,maxcpus=4  \
-numa cpu,node-id=0,socket-id=0 \
-numa cpu,node-id=0,socket-id=1 \
-numa cpu,node-id=1,socket-id=2 \
-numa cpu,node-id=1,socket-id=3 \
-numa dist,src=0,dst=1,val=20 \
-net nic \
-net user \
-hda testing.img \
-numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=40 \
-numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
-numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=80 \
-numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
-numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-latency,latency=80 \
-numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
-numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-latency,latency=40 \
-numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \

Then the latencies and bandwidths between the nodes were tested using the Intel Memory Latency Checker v3.9 (https://software.intel.com/content/www/us/en/develop/articles/intelr-memory-latency-checker.html). But the obtained results did not match the configuration. The following are the results obtained. 

Latency_matrix with idle latencies (in ns)

Numa
Node  0     1
0    36.2 36.4
1    34.9 35.4

Bandwidth_matrix with memory bandwidths (in MB/s)

Numa
Node 0       1
0 15167.1 15308.9
1 15226.0 15234.0

A test was also conducted with the tool “lat_mem_rd” from lmbench to measure the memory read latencies. This also gave results which did not match the config.

Any information on why the config latency and bandwidth values are not applied, would be appreciated.

On Sat, 25 Jul 2020 07:26:38 -0000
Vishnu Dixit <email address hidden> wrote:

> Public bug reported:
> 
> I was trying to configure latencies and bandwidths between nodes in a
> NUMA emulation using QEMU 5.0.0.
> 
> Host : Ubuntu 20.04 64 bit
> Guest : Ubuntu 18.04 64 bit
> 
> The machine configured has 2 nodes. Each node has 2 CPUs and has been
> allocated 3GB of memory. The memory access latencies and bandwidths for
> a local access (i.e from initiator 0 to target 0, and from initiator 1
> to target 1) are set as 40ns and 10GB/s respectively. The memory access
> latencies and bandwidths for a remote access (i.e from initiator 1 to
> target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s
> respectively.
> 
> The command line launch is as follows.
> 
> sudo x86_64-softmmu/qemu-system-x86_64  \
> -machine hmat=on \
> -boot c \
> -enable-kvm \
> -m 6G,slots=2,maxmem=7G \
> -object memory-backend-ram,size=3G,id=m0 \
> -object memory-backend-ram,size=3G,id=m1 \
> -numa node,nodeid=0,memdev=m0 \
> -numa node,nodeid=1,memdev=m1 \
> -smp 4,sockets=4,maxcpus=4  \
> -numa cpu,node-id=0,socket-id=0 \
> -numa cpu,node-id=0,socket-id=1 \
> -numa cpu,node-id=1,socket-id=2 \
> -numa cpu,node-id=1,socket-id=3 \
> -numa dist,src=0,dst=1,val=20 \
> -net nic \
> -net user \
> -hda testing.img \
> -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=40 \
> -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
> -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=80 \
> -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
> -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-latency,latency=80 \
> -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
> -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-latency,latency=40 \
> -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
> 
> Then the latencies and bandwidths between the nodes were tested using
> the Intel Memory Latency Checker v3.9
> (https://software.intel.com/content/www/us/en/develop/articles/intelr-
> memory-latency-checker.html). But the obtained results did not match the
> configuration. The following are the results obtained.
> 
> Latency_matrix with idle latencies (in ns)
> 
> Numa Node
> . .0. . .1.
> 0 36.2 36.4
> 1 34.9 35.4
> 
> Bandwidth_matrix with memory bandwidths (in MB/s)
> 
> Numa Node
> . . .0. . . .1. 
> 0 15167.1 15308.9
> 1 15226.0 15234.0
> 
> A test was also conducted with the tool “lat_mem_rd” from lmbench to
> measure the memory read latencies. This also gave results which did not
> match the config.
> 
> Any information on why the config latency and bandwidth values are not
> applied, would be appreciated.

There is no information about host hardware, so I'd onldy hazard a guess
that host is non NUMA machine so all guest RAM and CPUs are in the same
latency domain, so that's why you are seeing pretty much the same timings.

QEMU nor KMV do not simullate HW latencies at all, all that is configured
with '-numa hmat-lb' is intended for guest OS consumption as a hint for
smarter memory allocation and it's on to user to pin CPUs and RAM to
concrete host NUMA nodes and use  host's values in '-numa hmat-lb' to
actually get performance benefits from it on 'NUMA' machine.
On non NUMA host it's rather pointless except of the cases where one
needs to fake NUMA config (like testing some aspects of NUMA related code
in guest OS).

> 
> ** Affects: qemu
>      Importance: Undecided
>          Status: New
> 
> 
> ** Tags: bandwidth hmat hmat-lb latency
> 
> ** Description changed:
> 
>   I was trying to configure latencies and bandwidths between nodes in a
>   NUMA emulation using QEMU 5.0.0.
>   
>   Host : Ubuntu 20.04 64 bit
>   Guest : Ubuntu 18.04 64 bit
> -  
> - The machine configured has 2 nodes. Each node has 2 CPUs and has been allocated 3GB of memory. The memory access latencies and bandwidths for a local access (i.e from initiator 0 to target 0, and from initiator 1 to target 1) are set as 40ns and 10GB/s respectively. The memory access latencies and bandwidths for a remote access (i.e from initiator 1 to target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s respectively. 
> + 
> + The machine configured has 2 nodes. Each node has 2 CPUs and has been
> + allocated 3GB of memory. The memory access latencies and bandwidths for
> + a local access (i.e from initiator 0 to target 0, and from initiator 1
> + to target 1) are set as 40ns and 10GB/s respectively. The memory access
> + latencies and bandwidths for a remote access (i.e from initiator 1 to
> + target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s
> + respectively.
>   
>   The command line launch is as follows.
>   
>   sudo x86_64-softmmu/qemu-system-x86_64  \
>   -machine hmat=on \
>   -boot c \
>   -enable-kvm \
>   -m 6G,slots=2,maxmem=7G \
>   -object memory-backend-ram,size=3G,id=m0 \
>   -object memory-backend-ram,size=3G,id=m1 \
>   -numa node,nodeid=0,memdev=m0 \
>   -numa node,nodeid=1,memdev=m1 \
>   -smp 4,sockets=4,maxcpus=4  \
>   -numa cpu,node-id=0,socket-id=0 \
>   -numa cpu,node-id=0,socket-id=1 \
>   -numa cpu,node-id=1,socket-id=2 \
>   -numa cpu,node-id=1,socket-id=3 \
>   -numa dist,src=0,dst=1,val=20 \
>   -net nic \
>   -net user \
>   -hda testing.img \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   
>   Then the latencies and bandwidths between the nodes were tested using
>   the Intel Memory Latency Checker v3.9
>   (https://software.intel.com/content/www/us/en/develop/articles/intelr-
>   memory-latency-checker.html). But the obtained results did not match the
>   configuration. The following are the results obtained.
>   
>   Latency_matrix with idle latencies (in ns)
>   
>   Numa
> - Node  0     1
> + lat node0 node1
>   0    36.2 36.4
>   1    34.9 35.4
>   
>   Bandwidth_matrix with memory bandwidths (in MB/s)
>   
> - Numa
> - Node 0       1
> + Numa Node 
> + bw node0 .bw node1
>   0 15167.1 15308.9
>   1 15226.0 15234.0
>   
>   A test was also conducted with the tool “lat_mem_rd” from lmbench to
>   measure the memory read latencies. This also gave results which did not
>   match the config.
>   
>   Any information on why the config latency and bandwidth values are not
>   applied, would be appreciated.
> 
> ** Description changed:
> 
>   I was trying to configure latencies and bandwidths between nodes in a
>   NUMA emulation using QEMU 5.0.0.
>   
>   Host : Ubuntu 20.04 64 bit
>   Guest : Ubuntu 18.04 64 bit
>   
>   The machine configured has 2 nodes. Each node has 2 CPUs and has been
>   allocated 3GB of memory. The memory access latencies and bandwidths for
>   a local access (i.e from initiator 0 to target 0, and from initiator 1
>   to target 1) are set as 40ns and 10GB/s respectively. The memory access
>   latencies and bandwidths for a remote access (i.e from initiator 1 to
>   target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s
>   respectively.
>   
>   The command line launch is as follows.
>   
>   sudo x86_64-softmmu/qemu-system-x86_64  \
>   -machine hmat=on \
>   -boot c \
>   -enable-kvm \
>   -m 6G,slots=2,maxmem=7G \
>   -object memory-backend-ram,size=3G,id=m0 \
>   -object memory-backend-ram,size=3G,id=m1 \
>   -numa node,nodeid=0,memdev=m0 \
>   -numa node,nodeid=1,memdev=m1 \
>   -smp 4,sockets=4,maxcpus=4  \
>   -numa cpu,node-id=0,socket-id=0 \
>   -numa cpu,node-id=0,socket-id=1 \
>   -numa cpu,node-id=1,socket-id=2 \
>   -numa cpu,node-id=1,socket-id=3 \
>   -numa dist,src=0,dst=1,val=20 \
>   -net nic \
>   -net user \
>   -hda testing.img \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   
>   Then the latencies and bandwidths between the nodes were tested using
>   the Intel Memory Latency Checker v3.9
>   (https://software.intel.com/content/www/us/en/develop/articles/intelr-
>   memory-latency-checker.html). But the obtained results did not match the
>   configuration. The following are the results obtained.
>   
>   Latency_matrix with idle latencies (in ns)
>   
> - Numa
> + Numa Node
>   lat node0 node1
>   0    36.2 36.4
>   1    34.9 35.4
>   
>   Bandwidth_matrix with memory bandwidths (in MB/s)
>   
> - Numa Node 
> - bw node0 .bw node1
> + Numa Node
> + bw node0  bw node1
>   0 15167.1 15308.9
>   1 15226.0 15234.0
>   
>   A test was also conducted with the tool “lat_mem_rd” from lmbench to
>   measure the memory read latencies. This also gave results which did not
>   match the config.
>   
>   Any information on why the config latency and bandwidth values are not
>   applied, would be appreciated.
> 
> ** Description changed:
> 
>   I was trying to configure latencies and bandwidths between nodes in a
>   NUMA emulation using QEMU 5.0.0.
>   
>   Host : Ubuntu 20.04 64 bit
>   Guest : Ubuntu 18.04 64 bit
>   
>   The machine configured has 2 nodes. Each node has 2 CPUs and has been
>   allocated 3GB of memory. The memory access latencies and bandwidths for
>   a local access (i.e from initiator 0 to target 0, and from initiator 1
>   to target 1) are set as 40ns and 10GB/s respectively. The memory access
>   latencies and bandwidths for a remote access (i.e from initiator 1 to
>   target 0, and from initiator 0 to target 1) are set as 80ns and 5GB/s
>   respectively.
>   
>   The command line launch is as follows.
>   
>   sudo x86_64-softmmu/qemu-system-x86_64  \
>   -machine hmat=on \
>   -boot c \
>   -enable-kvm \
>   -m 6G,slots=2,maxmem=7G \
>   -object memory-backend-ram,size=3G,id=m0 \
>   -object memory-backend-ram,size=3G,id=m1 \
>   -numa node,nodeid=0,memdev=m0 \
>   -numa node,nodeid=1,memdev=m1 \
>   -smp 4,sockets=4,maxcpus=4  \
>   -numa cpu,node-id=0,socket-id=0 \
>   -numa cpu,node-id=0,socket-id=1 \
>   -numa cpu,node-id=1,socket-id=2 \
>   -numa cpu,node-id=1,socket-id=3 \
>   -numa dist,src=0,dst=1,val=20 \
>   -net nic \
>   -net user \
>   -hda testing.img \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-latency,latency=80 \
>   -numa hmat-lb,initiator=1,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=5G \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-latency,latency=40 \
>   -numa hmat-lb,initiator=1,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=10G \
>   
>   Then the latencies and bandwidths between the nodes were tested using
>   the Intel Memory Latency Checker v3.9
>   (https://software.intel.com/content/www/us/en/develop/articles/intelr-
>   memory-latency-checker.html). But the obtained results did not match the
>   configuration. The following are the results obtained.
>   
>   Latency_matrix with idle latencies (in ns)
>   
>   Numa Node
> - lat node0 node1
> - 0    36.2 36.4
> - 1    34.9 35.4
> + . .0. . .1.
> + 0 36.2 36.4
> + 1 34.9 35.4
>   
>   Bandwidth_matrix with memory bandwidths (in MB/s)
>   
>   Numa Node
> - bw node0  bw node1
> + . . .0. . . .1. 
>   0 15167.1 15308.9
>   1 15226.0 15234.0
>   
>   A test was also conducted with the tool “lat_mem_rd” from lmbench to
>   measure the memory read latencies. This also gave results which did not
>   match the config.
>   
>   Any information on why the config latency and bandwidth values are not
>   applied, would be appreciated.
> 



It indeed was a non NUMA machine