CPU 100% with tornado 4.5.3?

We have built a app with tornado 4.5.3, along with pynsq 0.8.2(which was not compatible with tornado 5), running on CentOS 7.4 with default kernel 3.10.0-693.21.1.el7.x86_64.
CPU usage came up to 100% soon after app started, with ~100 connections.
We are using tornado's TCPServer class and IOStream.read_bytes(4096, partial=True) to implement our async tcp server, something like

        while True:
            try:
                buf = await self._stream.read_bytes(4096, partial=True)
            except:
                self.close()
                return None
            return buf

Many lines like read(370, 0x2e47d80, 11) = -1 EAGAIN (Resource temporarily unavailable) in strace result, and strace -c result was:

% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 37.95    0.892246           4    245309    243731 read
 35.95    0.845239           3    242459         2 ioctl
 19.15    0.450160           4    119424           epoll_wait
  4.58    0.107586          13      7996       590 futex
  0.46    0.010713           3      3263      1395 recvfrom
  0.43    0.010007           5      1908           sendto
  0.33    0.007788           3      2700           epoll_ctl
  0.30    0.006951           4      1638           write
  0.19    0.004412           3      1746           fstat
  0.17    0.004001           2      1711      1705 lseek
  0.10    0.002335           3       754       726 getpeername

A tornado issue?

Start by examining the TCPServer usage and the IOStream.read_bytes(4096, partial=True) call described in the report, then reproduce the behavior with the stated Tornado, pynsq, CentOS, and kernel versions. Compare the event-loop activity with the repeated EAGAIN reads and epoll_wait calls; done means determining whether Tornado is responsible and documenting a reproducible cause or required configuration change.