From idea to profit

hardware

RTX 3090 vs A100 tensorflow performance

06.10.2020 – 00:23

When I searched how to estimate GPU performance I found this answer on stackoverflow, which contains the following code:

import os
import sys
import tensorflow as tf
import time

n = 8192
dtype = tf.float32
with tf.device("/gpu"):
    matrix1 = tf.Variable(tf.ones((n, n), dtype=dtype))
    matrix2 = tf.Variable(tf.ones((n, n), dtype=dtype))
    product = tf.matmul(matrix1, matrix2)


# avoid optimizing away redundant nodes
config = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))
sess = tf.Session(config=config)

sess.run(tf.global_variables_initializer())
iters = 10

# pre-warming
sess.run(product.op)

start = time.time()
for i in range(iters):
  sess.run(product.op)
end = time.time()
ops = n**3 + (n-1)*n**2 # n^2*(n-1) additions, n^3 multiplications
elapsed = (end - start)
rate = iters*ops/elapsed/10**9
print('\n %d x %d matmul took: %.2f sec, %.2f G ops/sec' % (n, n,
                                                            elapsed/iters,
                                                            rate,))

import os

import sys

import tensorflow as tf

import time

n = 8192

dtype = tf.float32

with tf.device("/gpu"):

matrix1 = tf.Variable(tf.ones((n, n), dtype=dtype))

matrix2 = tf.Variable(tf.ones((n, n), dtype=dtype))

product = tf.matmul(matrix1, matrix2)

# avoid optimizing away redundant nodes

config = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))

sess = tf.Session(config=config)

sess.run(tf.global_variables_initializer())

iters = 10

# pre-warming

sess.run(product.op)

start = time.time()

for i in range(iters):

sess.run(product.op)

end = time.time()

ops = n**3 + (n-1)*n**2 # n^2*(n-1) additions, n^3 multiplications

elapsed = (end - start)

rate = iters*ops/elapsed/10**9

print('\n %d x %d matmul took: %.2f sec, %.2f G ops/sec' % (n, n,

elapsed/iters,

rate,))

After Nvidia released a bunch of new generation GPUs I wanted to compare their performance.
To measure fp16 performance dtype was changed to tf.float16.
To benchmark matrix multiplication in tensorflow 2 compatibility mode was used. It can be enabled by replacing

import tensorflow as tf

1	import tensorflow as tf

with

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

1 2	import tensorflow.compat.v1 as tf tf.disable_v2_behavior()

So final results for tensorflow 2.4.0 are in table:

GPU	fp32 performance	fp16 performance
RTX 2080	10877.23 G ops/sec	42471.64 G ops/sec
V100	14743.50 G ops/sec	89348.57 G ops/sec
RTX 3090	35958.73 G ops/sec	69669.73 G ops/sec
A100	79158.13 G ops/sec	232681.81 G ops/sec
RTX 4090	80802.89 G ops/sec	162852.21 G ops/sec

Yuriy Nazarov | Permalink | Comment (0)

How to disable NVIDIA card in X server

20.10.2018 – 22:16

I actively using CUDA capable cards from NVIDIA for machine learning applications.
When GPU is under heavy load by ml tasks it’s almost impossible to do anything in graphical interface because screen refreshes too slow.
So I switched to integrated video controller. It’s not solved the issue with low refresh rate right away because X server still using NVIDIA card.
After some trials and errors I found a quick and dirty solution: disable graphics card driver during boot to prevent X server to use a card.
To accomplish this you need to blacklist driver by adding a following line:

install nvidia /bin/false

1	install nvidia /bin/false

to /etc/modprobe.d/blacklist.conf
And don’t forget to update initramfs:

sudo update-initramfs -u

1	sudo update-initramfs -u

After system boot you can load driver manually by following command:

sudo insmod /lib/modules/`uname -r`/updates/dkms/nvidia.ko

1	sudo insmod /lib/modules/`uname -r`/updates/dkms/nvidia.ko

Yuriy Nazarov | Permalink | Comment (0)

AVR electronic hardware

Wireless Arduino thermometer with data logger

20.07.2016 – 03:34

There is a bunch of stuff in my closet, so it requires only time and appropriate mood to build some hardware device. And there is a summer around and weather is rather hot, so it’s not a surprising idea to build a thermometer.
DS1820 is a great thermometer and I started with it, but I have a spare HDPM01 barometer module with built-in thermometer, so let’s use it.
read more »

Yuriy Nazarov | Permalink | Comment (0)

Lego soul

Nice birthday gift

10.07.2016 – 17:10

It’s well known that I’m a Lego fan. So I’m very happy to add such a beautiful plane to my Lego collection.

Read more to see functions animation. read more »

Yuriy Nazarov | Permalink | Comment (0)

git linux

Space efficient source code storage

23.03.2016 – 23:28

During implementation of DistTest I faced with necessity of building a lot of different linux kernel versions. As a first solution I chose downloading archives from kernel.org for each used version. But I soon realized that about 1000 versions of sources with size 0.5-1GB each would consume a lot of disk space. It’s also impossible to build kernel with exact commit precision using this approach.
Set of base versions with corresponding patches can save disk space, but uses a lot of random I/O during applying patches, so it’s slow on HDD and consume finite rewrite resource of SSD. Temporary nature of sources leads to conclusion “use tmpfs”. But aufs offers much less RAM consuming method – store in RAM only diffs.

Yuriy Nazarov | Permalink | Comment (0)

CPU hardware

Single threaded performance got stuck

08.11.2015 – 22:46

I fell in love with md5 hash algorithm because it can detect some very interesting characteristics of system which I want to benchmark. Almost all computations which need to be performed during computation of md5 hash sum are lying in critical path. It means that it’s almost impossible to parallelize md5 hash sum computation. And I’m not talking about execution in multiple threads, but about instruction level parallelization(superscalar and vector computing). So this feature excluding any new modern tricks used in CPU cores(like out-of-order execution and specialized instruction sets) out of equation and makes it perfect single thread benchmark.
Let’s see some numbers:
Calculate md5(10GiB of zerroes) on i5-760(Turbo frequency: 3.33 GHz, launch date Q3’10)(with Ubuntu 14.04)

$ (dd if=/dev/zero bs=1M count=10k | md5sum >/dev/null) 2>&1 | tail -n1
10737418240 bytes (11 GB) copied, 22.8959 s, 469 MB/s

1 2	$ (dd if=/dev/zero bs=1M count=10k \| md5sum >/dev/null) 2>&1 \| tail -n1 10737418240 bytes (11 GB) copied, 22.8959 s, 469 MB/s

And then do the same on i7-6700(Turbo frequency: 4.0 GHz, launch date Q3’15)(with Ubuntu 15.10)

$ (dd if=/dev/zero bs=1M count=10k | md5sum >/dev/null) 2>&1 | tail -n1
10737418240 bytes (11 GB) copied, 17.325 s, 620 MB/s

1 2	$ (dd if=/dev/zero bs=1M count=10k \| md5sum >/dev/null) 2>&1 \| tail -n1 10737418240 bytes (11 GB) copied, 17.325 s, 620 MB/s

So we have 140 and 155 MB/s per GHz respectively. It is 10.7% performance boost after 5 years of CPU evolution. And it looks so frustrating.
p.s. Yep, I know that CPU now much smarter than 5 years ago and have rich set of specialized instruction sets(like AES-NI which is responsible for +2200% ghash calculation speed). But any software developer should be ready for that fact that unparallelizeable algorithms execution will not become faster for even a bit in near future.

Yuriy Nazarov | Permalink | Comment (0)

perl programming

Selfmade dyndns

10.01.2015 – 02:38

Dyn.com stopped providing free dyndns hosting some time ago. It were sad news.
However most necessary functionality can be implemented in 70 lines of code.
To dynamically update dns records we must find a way to:

get current ip
update DNS record
schedule script for execution

1. Getting current ip address

Many ISP use NAT for security and money saving reasons. Also NAS used in routers. So, to get your IP address you need to ask your IP address from service located outside your network.
For example internet.yandex.ru web page ask resource http://ipv4.internet.yandex.ru/api/v0/ip to determine IP address. So, let’s use it.

use constant IP_API_URI         => 'http://ipv4.internet.yandex.ru/api/v0/ip';
sub get_ip($) {
	my ($ua) = @_;
	my $response = get($ua, IP_API_URI);
	return substr $response, 1, length($response)-2; #clean out quotes
}

use constant IP_API_URI => 'http://ipv4.internet.yandex.ru/api/v0/ip';

sub get_ip($) {

my ($ua) = @_;

my $response = get($ua, IP_API_URI);

return substr $response, 1, length($response)-2; #clean out quotes

}

2. Updating DNS record

Yandex provide API for DNS records management for domains parked in yandex. ~~API reference should be located here but link is currently broken :-(.~~ UPD: API reference.
In two words you should send POST request with token in header and data in request’s body to https://pddimp.yandex.ru/api2/admin/dns/edit url.
It should looks like following snippet:

use constant DNS_API_URI_PREFIX => 'https://pddimp.yandex.ru/api2/admin/dns/';
sub post($$$%){
	my ($ua, $url, $data, %headers) = @_;
	my $req = HTTP::Request->new(POST => $url);
	$req->header(%headers) if keys %headers;
	$req->content_type('application/x-www-form-urlencoded');
	$req->content($data);
	my $res = $ua->request($req);
	return $res->content if $res->is_success;
	return undef;
}
sub set_domain_ip($$$$$$){
	my ($ua, $token, $domain, $subdomain, $record_id , $ip) = @_;
	return decode_json post($ua, DNS_API_URI_PREFIX."edit", "domain=$domain&record_id=$record_id&subdomain=$subdomain&ttl=14400&content=$ip", PddToken => $token);
}

use constant DNS_API_URI_PREFIX => 'https://pddimp.yandex.ru/api2/admin/dns/';

sub post($$$%){

my ($ua, $url, $data, %headers) = @_;

my $req = HTTP::Request->new(POST => $url);

$req->header(%headers) if keys %headers;

$req->content_type('application/x-www-form-urlencoded');

$req->content($data);

my $res = $ua->request($req);

return $res->content if $res->is_success;

return undef;

}

sub set_domain_ip($$$$$$){

my ($ua, $token, $domain, $subdomain, $record_id , $ip) = @_;

return decode_json post($ua, DNS_API_URI_PREFIX."edit", "domain=$domain&record_id=$record_id&subdomain=$subdomain&ttl=14400&content=$ip", PddToken => $token);

}

3. Scheduling dns records updating

Use the cron, Luke :-) I mean that adding following line to cronjobs would be a simplies way to schedule updating of our dns records:

* * * * * ( token=123 /home/user/yandex_dns.pl --domain=domain.com --subdomain=subdomain )

1	* * * * * ( token=123 /home/user/yandex_dns.pl --domain=domain.com --subdomain=subdomain )

That’s all. Complete script on github.

Yuriy Nazarov | Permalink | Comment (0)

CPU hardware power

Few words about power management

09.12.2014 – 22:54

Not so long time ago I’ve faced with problem: on the same linux distributive some machines use Inte Turbo Boost but some others didn’t.
So… To investigate this problem I’ve read enough article about power management and want to summarize key aspects below.
Holy Grail of power management is ACPI(Advanced Configuration and Power Interface). It describes sleep(Sx), processor(Cx) and performace(Px) states.
Performance states came to replace legacy throtling(Tx) states.

S5(“Soft-Off”) All hardware is in off state.
S4(“Suspend to disk”) S5 + bootloader can determine this state + WOL available
S3(“Suspend to RAM”) RAM state preserved as well as S3 capable devices
S2(“Standby”) Almost the same as S3
S1(“Power On Suspend/Stopgrant”) All power on, but Hard Drives and not S1 capable devices is in off state, CPU is stopped.
S0 The system is turned on. Cx states is S0 substates
- C3(“Sleep”) cache not preserved
- C2(“Stop-Clock”) all preserved but clock is off
- C1(“Halt”) all preserved but CPU do nothing
- C0 is operating state. Px states is C0 substates
  - Pn Minimum frequency
  - P1 Maximum base frequency
  - P0 TurboBoost enabled

So… About TurboBoost issues solution. It was just a BIOS bug(or feature? Who knows?) that doesn’t moved cpu to P0 state on some boards and does on another.

Yuriy Nazarov | Permalink | Comment (0)

programming

Buzz words collection

16.07.2014 – 21:41

My collection of programmers buzz words. I think it will be useful for beginner programmers and Software Engineering students. Will be actively replenished.

Principles

Defensive programming
KISS – Keep it simple, stupid
DRY – Don’t repeat yourself
YAGNI – You Ain’t Gonna Need It
ACID – Atomicity Consistency Isolation Durability
Inversion of control
Uniform Access Principle
Law of Demeter
SOA – Service-oriented architecture
GRASP – General Responsibility Assignment Software Patterns
SOLID
- Single responsibility principle
- Open/closed principle
- Liskov substitution principle
- Interface segregation principle
- Dependency inversion principle

Yuriy Artamonov | Permalink | Comment (0)

linux

Yet another security note. Hardlinks.

21.03.2014 – 03:08

Let’s imagine that you have a bunch of files(with default mode “rw-r–r–“) and you configured automatic or performed manual hardlink based backup of them.
Well…
Then you moved one of these files to “secured” folder, that have strict rights (“drwx——” for example).
Before adding some confidential information to this file it was good idea to change file permissions to more strict one. But it is not clear how important it is because no one but owner can access file located at “secured/file” when “secured” folder have “drwx——” rights.
Well… Let’s preserve old permissions if changing of them is not necessary.
But what about hardlink to file saved in usual folder? Oh yes, file located at “usual/file” still can be opened by everyone.
Conclusion:
– You must remember about all hardlinks of your files when you think about security.
– Creating hardlinks by inode and opening file by inode denied by security reasons.

Yuriy Nazarov | Permalink | Comment (0)

From idea to profit

RTX 3090 vs A100 tensorflow performance

How to disable NVIDIA card in X server

Wireless Arduino thermometer with data logger

Nice birthday gift

Space efficient source code storage

Single threaded performance got stuck

Selfmade dyndns

1. Getting current ip address

2. Updating DNS record

3. Scheduling dns records updating

Few words about power management

Buzz words collection

Principles

Yet another security note. Hardlinks.

Recent Posts

Meta

Categories

Archives